On this page
- How to read the evidence labels in this guide
- The AI Spark method: define, filter, test, score, choose
- Step 1: Define the job
- Step 2: Filter by constraints
- Privacy and security: questions to ask before you paste anything
- Step 3: Run a small test
- Step 4: Score with evidence, not guesses
- Step 5: Choose, write it down and set a re-check
- General assistant or specialised tool?
- Free or paid? A decision rule, not a verdict
- Three decision exercises
- An illustrative example: choosing a tool for one job
- Red flags when you read a tool's claims
- What this guide does not claim
- Where to go next
- Frequently asked questions
- Sources
Most AI tool decisions start the wrong way round. Someone sees a demo, signs up, and only later asks what the tool was supposed to do. The trial then proves very little: nobody wrote down what "good" meant, so every output looks impressive or disappointing depending on the day.
This guide gives a small business a repeatable way to decide which AI tool fits one specific job. It is a method, not a list of products. We call it define, filter, test, score, choose. It includes a requirements card, a scorecard in which every cell needs a source, three decision exercises built from tests we have already published, and a clearly fictional example.
What this guide is not. For what a practical set of tools can look like, see the AI tool stack for a one-person business. For writing the instructions you give an AI, see how to write better AI prompts. Ai Spark is not an independent testing lab, and every tool mention says whether we tested it.
Disclosure. Ai Spark builds most of its workflows in n8n, so we have more evidence on n8n than on its alternatives. We say so wherever it matters. There are no affiliate links in this guide.
Key takeaways
- Write the job and the checks before you look at any tool. A test whose checks were written after the output proves very little.
- Filter by constraints first (data, connections, human approval, usage limits), then test two or three candidates, not ten.
- Fill a scorecard only with evidence: a test, an official page or your own stated requirement. Otherwise write NOT TESTED or UNKNOWN.
- Free or paid is a decision rule, not a verdict. Move to paid when a verified limit blocks something you must have.
- We ran no new brand-versus-brand tests for this guide. The exercises use tests we had already published plus official pages, and they say what is still untested.
How to read the evidence labels in this guide
| Label | Meaning |
|---|---|
| TESTED BY AI SPARK | We ran it ourselves and recorded what happened, with the limits stated. |
| OFFICIAL SOURCE ONLY | It comes from the vendor's or publisher's own documentation, listed under Sources. We did not test it. |
| ILLUSTRATIVE EXAMPLE | Made up to explain an idea. Not a client, not a result. |
| OUR RECOMMENDATION | Our own judgement. It is not a test result. |
| NOT TESTED / UNKNOWN | NOT TESTED: we have not tried it, so we do not describe how it behaves. UNKNOWN: we have no evidence either way. |
Figure 1: The evidence legend. The border style repeats the label, so the meaning does not depend on colour.
Sources were checked on 30 October 2026 (Pakistan time, PKT). Plan limits and terms change, so check the current versions before you rely on them.
The AI Spark method: define, filter, test, score, choose
This is our own practical framework. It is not an industry standard and it is not taken from any vendor. It turns "which tool should I buy?" into five smaller questions that can each be answered with evidence.
Two official sources point the same way on the key idea. Anthropic's documentation says good success criteria are specific, measurable, achievable and relevant. OpenAI's evaluation guide says to start from a defined objective and to test on inputs that mirror real use, including unusual ones. OFFICIAL SOURCE ONLY. The five-step packaging below is ours.
Figure 2: The five steps. Step 3 is highlighted because the rest of this guide leans on it.
| Step | The question it answers | What you write down | A stop sign |
|---|---|---|---|
| 1. Define | What exactly must the tool do? | A one-sentence job, the input, the output and how a person will check the result. | You cannot say what a good result looks like. |
| 2. Filter | Which candidates are even allowed? | Your constraints (data, connections, approval, limits) and a shortlist of two or three. | No candidate can handle your data. |
| 3. Test | Does it do this job on our own inputs? | The checks, written before the run; a few inputs including one awkward one; the date, tool and plan. | The output varies and you ran it only once. |
| 4. Score | What does the evidence say, requirement by requirement? | Met, not met, not tested or unknown, each with a source. | A must-have is not met. |
| 5. Choose | Which one for now, and when do we look again? | The decision, the reasons and a re-check date. | The decision rests on an unknown must-have. |
Step 1: Define the job
Write one sentence in this shape: when this happens, the tool takes this input and produces this output, and this person checks it like this. If you cannot finish the sentence, you are not ready to compare tools.
Then answer ten questions. They are ours, and they are written so that each answer can later be tested or looked up.
| Requirement question | What to write down | Why it matters |
|---|---|---|
| 1. Job to be done | The task in one sentence, and how often it happens. | How often the task happens affects how much setup effort it can justify. |
| 2. Required input | What goes in (text, files, customer details) and where it comes from. | The input decides the privacy and connection questions below. |
| 3. Required output | Format, length, language, and where the result goes next. | A fixed format is something you can check automatically. |
| 4. Accuracy and review | What a wrong result would cost, and who checks it. | The higher the cost, the more a person must see before anything leaves the business. |
| 5. Connections | Which apps the tool must read from or write to. | A tool that connects to nothing creates copy-and-paste work. |
| 6. Data and privacy | What data is entered, and whether it includes customer or staff data. | Terms can differ by product and plan (see the privacy section). |
| 7. Automation | Whether it runs by itself, on a schedule, or only when a person asks. | Unattended runs need stricter checks. |
| 8. Human approval | What a person must approve before anything is sent, saved or changed. | Draft-first is easier to undo than send-first. |
| 9. Cost and usage limit | What you are willing to spend, and which usage limit would stop you. | A limit you reach in week two is a cost. |
| 10. Testability | Whether you can try it on safe, realistic examples before connecting real data. | If you cannot test it, you can only trust it. |
Does the job need AI at all? Add a baseline to every shortlist: a fixed template, a spreadsheet formula or a checklist. If the baseline meets every must-have, you may not need an AI tool for this job. If it does not, you now know exactly what the AI has to add. OUR RECOMMENDATION.
Step 2: Filter by constraints
Mark each requirement MUST (the option is out if it fails), SHOULD or NICE. Then remove any candidate that cannot meet a MUST on paper, and keep two or three. More than that makes testing slow, because every candidate needs evidence for every MUST. OUR RECOMMENDATION.
Where a fact should come from matters as much as the fact. A listicle or a comparison site is not a source for a product's data terms or limits.
| If you need to know | Look here | Why not a listicle |
|---|---|---|
| Whether it connects to your apps | The vendor's integrations or documentation page, then your own test. | Such pages may not say which plan, version or date a claim applies to. |
| How your data is handled | The terms or data-controls page for the exact product and plan you will use. | The answer can differ between a chat app, an API, a free tier and a paid tier. |
| Usage limits and what a free plan includes | The vendor's current plan or pricing page, with the date you read it. | Limits change, and a copied table is already out of date. |
| Whether a person can approve a step | The vendor's documentation for that feature, then your own test. | A feature name does not show how approval works until you read the documentation and try it. |
| Whether it does the job well | Your own test on your own inputs. | Nobody else's examples are your examples. |
Privacy and security: questions to ask before you paste anything
This section is practical due diligence, not legal advice. Laws and obligations differ by place and by situation. Ask these before any real customer or staff data goes into a tool:
- What data will be entered, and does it include customer or staff data?
- What do the terms for this exact plan say about using prompts and outputs, retention and human review?
- Who in your business can see what goes in and what comes out?
- Does the tool need to sign in to your accounts (email, sheets, CRM)? What access does it ask for?
- Can that access be limited, and can you remove it later?
- Does a person approve anything before it is sent or saved?
Three official examples show why the "exact plan" question matters. They are examples, not a ranking, and none of them is a statement about every product from that company.
Terms can differ by service category, and by region. Google's Gemini API terms treat "unpaid" and "paid" services differently. For unpaid services, the terms say Google may use prompts and responses to improve its products and that human reviewers may read them, and they advise against submitting sensitive, confidential or personal information. For paid services, they say Google does not use prompts or responses to improve its products. But the terms also say that in the European Economic Area, Switzerland and the United Kingdom the paid-service data terms apply to all services, including free ones. Whether a service counts as paid can depend on the product and on whether billing is active. We are not saying that any one plan is safe or unsafe. The point is that you must check which service category applies to you, whether billing is active, your region and the current terms before you paste anything. OFFICIAL SOURCE ONLY, checked 30 Oct 2026.
API terms are not the same as chat-app terms. OpenAI's data-controls page says that, by default, data sent to its API is not used to train or improve its models unless the customer opts in. It says abuse-monitoring logs are generally kept for up to 30 days by default, and may be kept longer where the law requires or where needed to protect the service or others. Retention and stored application state also vary by endpoint and feature. Zero Data Retention is available only to customers OpenAI approves, so it is not something every account has by default. OFFICIAL SOURCE ONLY, checked 30 Oct 2026. Read the page for the product and feature you will actually use.
Access you grant can be reviewed. Google Account Help explains that you can see which third-party apps have access to your Google Account and remove that access, and that an app may still keep data it already collected. OFFICIAL SOURCE ONLY, checked 30 Oct 2026. Our own advice is to connect with the least access that works, to keep a note of every connection you make and to remove access when you stop using a tool. OUR RECOMMENDATION. In our own tests, a saved Google connection expired once and had to be reconnected. TESTED BY AI SPARK (n8n, 18 Oct 2026).
Human approval can be built in. n8n's documentation describes a human-in-the-loop option for AI Agent tool calls. OFFICIAL SOURCE ONLY, checked 30 Oct 2026. It says a person can approve or deny selected AI Agent tool calls before they run, in channels such as Chat, Slack, Gmail or Microsoft Teams. AI Spark did not test this feature. What we did test was simpler: our workflow created a Gmail draft and sent nothing (TESTED BY AI SPARK, run #2689, 18 Oct 2026; see the lead follow-up tutorial).
Until you have read the terms for your plan, test with invented or public examples. That is what we do. OUR RECOMMENDATION.
Step 3: Run a small test
A small test is not a trial period. It is a short, planned check with the pass marks written down first.
- Write the checks before you run anything. For example: "the output is in this format", "it does not state a policy I did not supply", "nothing is sent". Anthropic's guidance on evaluations makes the same point: define success criteria first. OFFICIAL SOURCE ONLY.
- Pick three to five inputs, and make one of them awkward. Use a missing detail, an off-topic message or a message that tells the tool to ignore its instructions. OpenAI's evaluation guide lists the same kinds of edge cases. OFFICIAL SOURCE ONLY.
- Run it, and run it again if the output can vary. OpenAI's guide says generative AI is variable, which is why single results can mislead. OFFICIAL SOURCE ONLY.
- Keep it safe. Use a draft, a test account or a copy of a sheet. Send nothing.
- Record what happened. The date, the tool, the plan or version, what you entered, what came out, and which checks failed.
What our own small tests showed. Three examples, all small and all with limits:
- In one controlled single-run test (30 Oct 2026 PKT) a Claude draft kept to a required format and did not state a refund policy we had not supplied. The exact model was not independently verified, and one run does not show it will happen again. TESTED BY AI SPARK. See AI customer support triage and human handoff.
- In an earlier single-run prompt test, the basic prompt answered "Yes, we offer monthly bookkeeping" without being told what the business offered. TESTED BY AI SPARK, dated 29 October 2026 in that guide. The same tool, a different instruction, a different result: the instruction mattered as much as the tool.
- In our n8n build, the first run failed at the Google Sheets step, and our login session expired partway through setup. TESTED BY AI SPARK. A feature list would not have shown either problem. A test did.
None of these is a benchmark, a blind test or a statistical result, and none was run on a production system. What a small test cannot tell you: how a tool behaves at volume, over months, after a vendor update or with a support request. Say so in your decision record.
Step 4: Score with evidence, not guesses
A scorecard is only as honest as its cells. First, set the weights, and set them before you look at any tool:
- MUST: the option is out if it fails this requirement.
- SHOULD: important, but you could live without it if the rest fits.
- NICE: a bonus that should not decide the choice on its own.
Then fill the cells with these rules:
- A cell may say MET or NOT MET only if it comes from your own test (T), an official page you have read and dated (O) or a requirement you have set yourself (R).
- Anything else is NOT TESTED, UNKNOWN or NOT APPLICABLE. Never fill a gap with a guess to make the table look complete.
- Set the weights before you score: MUST, SHOULD or NICE. A MUST that is not met removes the option. A MUST that is not tested sends you back to Step 3.
- UNKNOWN is not zero, NOT TESTED is not a failure and NOT APPLICABLE is not a failure. They mean the evidence is missing or the requirement does not apply. Only NOT MET counts against an option.
- If you want a number, write the formula first, count only what is met, and show the unknown cells next to the total. A total built on guesses is worse than no total. OUR RECOMMENDATION.
Figure 3: The layout of an evidence scorecard. The values are placeholders, not real products. T = tested by us, O = official source, N/A = not applicable.
Here is a blank version you can copy:
| Requirement | Level | Option A | Option B | Option C | Evidence source | Notes |
|---|---|---|---|---|---|---|
| (a requirement from Step 1) | MUST / SHOULD / NICE | MET / NOT MET / NOT TESTED / UNKNOWN / N/A | T (your test), O (page and date), R (your requirement) | |||
TEMPLATE. It contains no data. Every filled cell should show its source (T, O or R).
Step 5: Choose, write it down and set a re-check
- Every MUST is met. The option passes. If two options pass, prefer the one whose evidence you trust more and whose access and cost you can limit.
- A MUST is not met. The option is out, or you change the requirement and write down why.
- A MUST is not tested or unknown. Run the next test before you pay or connect live data.
- No option passes. Keeping the manual process is a valid decision.
Keep a short decision record: the job, the choice, the evidence, what was not tested, the date and a re-check trigger. Good triggers are a change to the plan or its terms, a new model or version, a change in your requirements and any failed check. OUR RECOMMENDATION. There is no fixed schedule that we have evidence for, so tie the re-check to an event rather than a calendar.
General assistant or specialised tool?
Neither is better in general. They do different jobs, and one workflow can use both. For a full stack design, see the AI tool stack guide; here we only need the difference.
| General AI assistant | Specialised business tool (workflow, CRM, helpdesk) | |
|---|---|---|
| Typical job | Drafting, summarising, sorting and planning from text you paste in. | Running steps in a set order, storing records, scheduling and moving data between apps. |
| Check first (our recommendation) | It can state things that were not supplied, and its format can vary with the instructions; see our single-run prompt tests. | Connection limits, run limits, setup effort and what happens when a step fails. |
| Our evidence | TESTED BY AI SPARK: single-run Claude tests for drafting and sorting. One assistant only; no other assistant tested. | TESTED BY AI SPARK: a same-task Google Sheets test in n8n, Zapier Free and Make Free, and an n8n form, sheet and Gmail draft workflow. CRMs: OFFICIAL SOURCE ONLY in our sales pipeline guide; not tested. |
A workflow can combine them: the automation platform moves the data and the assistant writes a draft. n8n's documentation describes asking a person to approve or deny an AI Agent's tool call. OFFICIAL SOURCE ONLY. The workflow we tested had no AI step, so we have not tested the combination. TESTED BY AI SPARK for the parts we ran; NOT TESTED for the AI step.
Free or paid? A decision rule, not a verdict
We do not say "free is enough" or "paid is better". Both depend on your must-haves.
- Stay on free when your test and every must-have fit inside the current limit and the terms suit your data.
- Consider paid when a verified limit blocks an important requirement.
- Do not pay for "we might need it later". A limit you have not reached is not a reason.
Three examples of verified limits, none of which is a recommendation, and none quotes a price:
- Steps. At the time we checked on 30 Oct 2026, Zapier's Free plan documented two-step Zaps: one trigger and one action. A two-step flow fits that structure. A flow with a form trigger, a Google Sheets step and a Gmail draft step needs more than two steps, so it would not fit the Free structure as documented. OFFICIAL SOURCE ONLY, checked 30 Oct 2026; plan limits can change. Our own two-step test needed nothing paid. TESTED BY AI SPARK.
- Data terms. As above, Google's Gemini API terms treat unpaid and paid services differently, with a regional exception. If your must-have is that your data is not used to improve a product, check the service category, billing status and region before you decide. OFFICIAL SOURCE ONLY, checked 30 Oct 2026.
- Who operates it. n8n's documentation describes n8n Cloud as the hosted option, and self-hosting as running n8n on your own infrastructure, where you provide and maintain that environment. OFFICIAL SOURCE ONLY, checked 30 Oct 2026. That is an owner responsibility and a cost to weigh against your must-haves, not a verdict for or against either option. We self-hosted Community Edition in our tests and did not test n8n Cloud.
Plans and limits change. Date every limit you rely on.
Three decision exercises
These are not contests between brands. Each one started from written requirements, with the evidence rules fixed before we filled in any cell. We did not run new brand-versus-brand tests for this guide, so we used tests we had already published and official pages. Every NOT TESTED below is a real gap.
| Exercise | What the evidence supports for now | Biggest gap |
|---|---|---|
| A. Draft a structured reply to a customer enquiry | A general AI assistant, kept to drafts, alongside a plain template as the baseline. | No other assistant or helpdesk feature was tested, and nobody has read the data terms for your plan. |
| B. Move a form enquiry into a sheet and a Gmail draft | n8n currently has the strongest evidence in AI Spark's own records. This is not a ranking of the products. Make Free is the next to test. | The Gmail draft and a form trigger were not tested in Zapier or Make, and we know n8n best. |
| C. Keep track of a small lead pipeline | Example decision: start with a spreadsheet and a status column, if your own requirements look like these. Not a rule for every business. | No CRM was tested, and reminders and exports were not tested in any tool. |
Open each record to see the requirements, the scorecard and the reasoning.
Exercise A: draft a structured reply to a customer enquiry
Task. Turn a pasted customer enquiry into a draft reply that a person reviews.
Must-have. A1 accepts pasted text plus written instructions. A2 returns a fixed format. A3 does not state facts or policies that were not supplied. A4 stays a draft; the tool sends nothing. A5 data handling is acceptable for the data you paste. Nice-to-have. Saved instructions; machine-readable output; can be called from an automation.
Options considered. A general AI assistant; an AI feature inside a helpdesk or CRM; a fixed email template with no AI as the baseline.
| Requirement | Level | General AI assistant | AI feature in a helpdesk or CRM | Source |
|---|---|---|---|---|
| A1 pasted text and instructions | MUST | MET in one run (Claude) | NOT TESTED | T: support guide, Test B |
| A2 fixed format | MUST | MET in one run | NOT TESTED | T: support guide, Test B |
| A3 no invented policy | MUST | MET in one run when told not to state a policy. A basic prompt in another run invented a service. | NOT TESTED | T: support guide Test B; prompt guide Test 1 |
| A4 stays a draft | MUST | MET: we copied the text by hand and nothing was sent | NOT TESTED | T: support guide |
| A5 data handling fits your data | MUST | UNKNOWN until you read the terms for your plan; terms differ by vendor and plan | UNKNOWN | O: Gemini API terms and OpenAI data controls, checked 30 Oct 2026 (examples) |
| Saved instructions | NICE | NOT TESTED | NOT TESTED |
What was not tested. Any other assistant, any helpdesk AI feature, repeated runs, real customer data and languages other than English.
Decision for now. Shortlist the general AI assistant, used for drafts only, and keep a fixed template as the baseline for questions that never change. The helpdesk feature is not shortlisted yet, and that is because we have no evidence either way, not because it failed. Why. The assistant is the only option with evidence on A1 to A4, but that is one assistant, one run and invented messages. It is a place to start, not a verdict. Next step: read the terms for your plan (A5), then repeat the same test on your own invented examples.
Exercise B: move a form enquiry into a sheet and a Gmail draft
Task. A form enquiry becomes a row in a Google Sheet and a Gmail draft reply, with nothing sent.
Must-have. B1 starts from a form or test input. B2 appends a row to a Google Sheet. B3 creates a Gmail draft, not a sent email. B4 shows what ran and whether it worked. B5 the full flow fits the plan you will use. Nice-to-have. A self-hosting option; an AI step later.
Options considered. n8n (self-hosted Community Edition, v2.33.6), the Zapier Free plan and the Make Free plan.
| Requirement | Level | n8n (self-hosted) | Zapier Free | Make Free | Source |
|---|---|---|---|---|---|
| B1 form or test input | MUST | MET (form trigger, run #2689) | NOT TESTED (we used a schedule trigger) | NOT TESTED (we started with "Run once") | T: lead follow-up tutorial; Google Sheets test |
| B2 append a Google Sheets row | MUST | MET, one run | MET, one editor test | MET, one run | T: Google Sheets test, 18 Oct 2026 |
| B3 Gmail draft, not send | MUST | MET (run #2689; Send not used) | NOT TESTED | NOT TESTED | T: lead follow-up tutorial |
| B4 shows what ran | MUST | MET (Executions list) | NOT TESTED for a published Zap (ours was a draft, so no history appeared) | MET (History tab) | T: Google Sheets test |
| B5 three-step flow fits the plan | MUST | MET in our run; we met no plan limit (self-hosted Community v2.33.6) | Not possible on the Free structure as documented on 30 Oct 2026 (one trigger plus one action) | NOT TESTED (three-step flow not run) | T for n8n; O: Zapier pricing page, 30 Oct 2026 |
| Self-hosting option | NICE | MET (documented) | UNKNOWN (not reviewed) | UNKNOWN (not reviewed) | O: n8n docs |
| AI step later | NICE | NOT TESTED (our workflow had no AI step) | NOT TESTED | NOT TESTED |
What was not tested. The Gmail draft step and a form trigger in Zapier or Make; a published Zap's history; volume; reliability over time; paid plans; n8n Cloud; any AI step. The Google Sheets test was one run in each tool and is not a benchmark.
Best product, or best-supported choice? These are different questions. We cannot say which product is best. We can say which choice is best supported by our current evidence. For this exact flow, n8n currently has the strongest evidence for these requirements in AI Spark's own records, because it is the only option we ran from end to end. Zapier Free, as documented on 30 Oct 2026, does not fit a three-step flow, so it stays on the list only if a paid plan is acceptable. Make Free is untested for the form and the Gmail draft, so it is the next test. We rank neither Make nor Zapier on tests we did not perform. A bias to know about. Ai Spark builds most of its workflows in n8n, so we have more evidence for it because we have used it more. More evidence is not the same as a better tool.
Exercise C: keep track of a small lead pipeline
Task. Keep a small list of leads with a stage and a next step.
Must-have. C1 each lead has a stage or status. C2 you can see which leads are open and what the next step is. C3 the data can be exported. C4 you can change it without a developer. C5 follow-up stays with a person. Nice-to-have. Reminders; several users; automation triggers.
Options considered. A spreadsheet (Google Sheets) and a CRM. HubSpot, Zoho CRM and HighLevel appear only as examples of the category.
| Requirement | Level | Spreadsheet | CRM (category) | Source |
|---|---|---|---|---|
| C1 stage or status per lead | MUST | MET: our workflow wrote a lead row and set its Status to "Draft Prepared" | Documented in the official pages we read on 26 Oct 2026 (examples only); not tested | T: run #2689; O: documentation read 26 Oct 2026 in our sales pipeline guide |
| C2 see open leads and next step | MUST | NOT TESTED (we built no filtered view) | NOT TESTED | No evidence |
| C3 export | MUST | NOT TESTED | UNKNOWN (not read) | No evidence |
| C4 change without a developer | MUST | MET: we set up the sheet columns by hand | UNKNOWN | T |
| C5 follow-up stays with a person | MUST | MET by design: the workflow only created a draft | NOT TESTED | T: run #2689 |
| Reminders, several users, triggers | NICE | NOT TESTED | NOT TESTED | No evidence |
What was not tested. Any CRM in a published test; reminders; several users; exports from any tool.
Example decision. OUR RECOMMENDATION. Given the requirements above, which ask for little beyond a status and a next step, a spreadsheet is where we have evidence, so we would start there and keep checking C1 to C5. This is not a rule that every small business should avoid a CRM at first. If your requirements include several people sharing one list, reminders or automation triggers, a CRM may be the right starting point, and you would run the same small test on it after checking what the vendor currently offers. Why. The decision follows your written requirements, not the tool category. See the sales pipeline guide and CRM automation workflows.
An illustrative example: choosing a tool for one job
ILLUSTRATIVE EXAMPLE — NOT A CLIENT CASE STUDY. Northfield Tutoring is a made-up one-person tutoring business. The details, options and outcomes are invented to show the method. Option A, B and C are placeholders, not real products. It is not a test, and it shows no time saved, revenue, conversions or customer results.
| Step | What happens in the example |
|---|---|
| Problem | Tutoring enquiries arrive by email. The owner wants each one saved as a row in a sheet and a draft reply she can edit before sending. |
| Requirements (Step 1) | MUST: the enquiry becomes a row with name, subject and year group; the draft uses only approved wording; the tool sends nothing; no pupil names go into a tool whose data terms she has not read. SHOULD: it runs without her opening anything. NICE: a weekly list of open enquiries. |
| Filter (Step 2) | Three candidates plus a baseline. Option A: an AI assistant used by hand. Option B: an automation platform with an AI step. Option C: an email template with merge fields as the baseline. A fourth option is dropped because no data terms could be found for it. |
| Test (Step 3) | Ten invented enquiries, one with no subject and one telling the tool to ignore its instructions. The checks were written first. Each option is run once. |
| Scorecard (Step 4) | Option A: all MUSTs met, but she must copy and paste. Option B: format met; the year group was missed on one invented enquiry; data terms UNKNOWN. Option C: meets the draft and format MUSTs but cannot read an enquiry, so it fails the row MUST. |
| Decision (Step 5) | Not yet. Option B has an unknown MUST, so she cannot choose it. She keeps using Option A by hand with invented samples, reads Option B's data terms, and re-runs the test on B only after that. She writes the date and the re-check trigger in her decision record. |
The point of the example is not which option wins. It is that an unknown MUST stops the decision, and that the record says what happens next.
Red flags when you read a tool's claims
- No data terms you can find for your plan.
- A claim you cannot check, such as "10x your results", with no worked example you could repeat.
- No way to export your data, templates or content.
- A free plan with no stated limits, or limits that appear only after you sign up.
- Features described only in marketing language and not in documentation.
These are red flags to look into, not proof of a bad tool. OUR RECOMMENDATION.
What this guide does not claim
| We do not claim | Because |
|---|---|
| That any tool is best or worst | We tested only some tools, on small tasks, in our own accounts. |
| That any tool saves time or money | We measured no time, cost or productivity. |
| That free is enough, or that paid is better | It depends on your must-haves and on limits that change. |
| That results will repeat | Our tests were single runs. Outputs can differ on another run. |
| That documentation is our testing | Every OFFICIAL SOURCE ONLY item comes from the vendor's pages and is labelled. |
| That this is legal or compliance advice | It is not. Laws and responsibilities vary by place and by situation. |
Where to go next
Want help applying this to a real workflow? Our automation audit guide shows how we map a process before choosing any tool. Ai Spark is a small automation studio, not an independent testing lab. If you would like a second pair of eyes on your requirements and a plan for a small test, get in touch.
- AI tool stack for a one-person business: what a practical set of tools can look like.
- How to write better AI prompts: how to instruct an AI and check what comes back.
- n8n vs Zapier vs Make: the same-task test behind Exercise B.
Frequently asked questions
How many tools should I compare?
Two or three, plus a baseline that uses no AI. Each extra candidate needs evidence for every must-have, so a long list mostly produces unknown cells. OUR RECOMMENDATION.
Are free AI tools good enough?
It depends on your must-haves. Stay on free while your test and your must-haves fit the current limit and the data terms suit you. Consider paid when a verified limit blocks something important. See the free-or-paid section above.
Can I just ask an AI assistant which tool to choose?
You can use it to brainstorm requirement questions or candidates. Its recommendations are not evidence. Check any claim against the vendor's own documentation and against your own small test. OUR RECOMMENDATION.
How often should I re-check a decision?
When something changes: the plan or its terms, the tool's version or model, your requirements, or a check that starts failing. We do not know of a fixed schedule that is right for everyone, so tie the re-check to an event and date your decision record.
Sources
All checked on 30 October 2026.
- Claude Platform Docs: Define success criteria and build evaluations (checked 30 Oct 2026)
- OpenAI API Docs: Evaluation best practices (checked 30 Oct 2026)
- OpenAI API Docs: Data controls in the OpenAI platform (checked 30 Oct 2026)
- Google AI for Developers: Gemini API Additional Terms of Service (checked 30 Oct 2026)
- Zapier: Pricing (Free plan limits, checked 30 Oct 2026)
- n8n Docs: Choose how to use n8n (checked 30 Oct 2026)
- n8n Docs: Human-in-the-loop for tools (checked 30 Oct 2026)
- Google Account Help: Share some access to your Google Account data with apps from other developers (checked 30 Oct 2026)
Our own tests are in n8n vs Zapier vs Make, the lead follow-up tutorial, the prompt guide and the support triage guide. Ai Spark is not affiliated with Anthropic, OpenAI, Google, n8n or Zapier, and we do not rank or endorse products. Product names appear only to point to documented features or to say which tool we tested. Last reviewed 30 October 2026.