What AI Implementation Actually Involves for a Small Business

For a small business, AI implementation is six pieces of work: choosing one job and measuring it, picking the tool (often one you already have), connecting it to your data with the right permissions, writing and testing instructions on real cases, rolling it out with a review step, and handing it to an owner who maintains it.

It rarely involves training a model, hiring a data scientist or buying hardware. As a rough planning range, a first workflow takes 20 to 60 hours of internal time spread over four to eight weeks, and most of those hours go on unglamorous parts: agreeing what "done" means, tidying up access, and testing the awkward cases that don't fit the pattern.

Follow me on Instagram@sagnikteaches

The six pieces of work, and who usually does each

The hours below are planning estimates for one workflow in a firm of under twenty people. They vary with how tidy your systems are, but the proportions hold: testing and rollout take longer than setting up the tool.

Connect on LinkedInSagnik Bhattacharya

Piece 1 is the one most often skipped, and skipping it makes the whole project unprovable. If nobody recorded that replying to an enquiry took 25 minutes, nobody can say whether 10 minutes is a win. Setting a baseline before you introduce AI covers the logging in detail.

Subscribe on YouTube@codingliquids

Piece 2 should start with what you already own. Microsoft 365 business plans include Copilot Chat, and Google Workspace business plans have Gemini built in, so the right first step can be switching on something your plan already includes. Whether a given app will talk to an AI tool is covered in whether AI can work with the tools you already use.

A quick version of piece 2 at an illustrative three-person estate agency that wants help writing property listings. Option one is Gemini, already in its Google Workspace plan, at no extra cost. Option two is ChatGPT Business, which needs at least two seats at $25 each billed monthly, so $50 a month. Option three is a specialist listings tool with its own subscription and its own login to manage. The agency runs five recent listings through Gemini first, pasting in the same viewing notes each negotiator used. Four drafts are usable after a light edit, and the fifth, a flat with an awkward lease, needed rewriting by hand whichever tool wrote it. Good enough: the decision takes an afternoon and costs nothing, and the other two options stay on the list in case volumes grow.

What isn't part of the work for most small firms

Owners often picture implementation as something much bigger than it is. For most first workflows, these are not involved:

• Training a model. The vendor has already trained it. You supply context: your instructions, a few examples, the documents it should answer from.
• A data project before you start. You need the data for this one job to be findable and reasonably consistent, not the whole business "AI-ready".
• New hardware or servers. Everything runs in the vendor's cloud.
• Custom code. Automation platforms connect most common business apps without it.

Custom development does come in when an app has no ready-made connection, when volumes make per-task pricing expensive, or when the logic is too branching for a drag-and-drop builder. That's usually a second or third workflow decision, not a first one.

The three parts owners underestimate

The cases that don't fit

Every process has a share of oddities: the enquiry that's really a job application, the invoice in the wrong currency format, the client who replies to a three-month-old thread. When you test on past cases, deliberately include the strange ones. A workflow that handles 85% of cases well and quietly mangles the other 15% is worse than no workflow, because people stop checking. Decide in advance what happens to anything the AI can't classify with confidence: it should land in front of a person, labelled "unclear".

A test log only needs a few columns. Here are four rows from a plumbing and heating firm testing supplier-invoice capture against past invoices (illustrative):

Case | What it was                          | Expected          | AI result          | Pass?
07   | Standard merchant invoice            | $1,284.50, #A7731 | $1,284.50, #A7731  | Yes
12   | Credit note laid out like an invoice | Credit, -$96.00   | Invoice, $96.00    | NO
19   | Two invoices scanned into one PDF    | Two bills         | One bill, 1st only | NO
23   | Handwritten total on a delivery note | Unclear           | $310 (a guess)     | NO

Rows 12 and 19 are the ones that would have done real damage in live use: a credit note posted as a bill means paying a supplier who owes you money. Both become explicit instructions ("if the document says credit note or shows a minus total, label it credit"; "if a PDF contains more than one invoice number, mark it unclear"), and row 23 confirms that anything handwritten goes to a person.

Access and account admin

Connecting an automation tool to email usually means granting it permission to read and send from a mailbox. Decide which mailbox, and connect a shared business one rather than someone's personal inbox. Make sure the automation account is registered to the company, paid on a company card, and has at least two people who can log in. This is dull work and it's where most ownership problems start.

A realistic version of how it goes wrong: at a small lettings agency, the office administrator sets up an automation under her own login, on a free trial she later upgrades on her own card because it was quicker than asking. It runs well for a year. When she leaves, her email account is closed, the automation account's password resets now go to a dead inbox, and her card is cancelled. The workflow stops at the next billing date, and nobody at the agency can log in to see what it did or switch it back on. Rebuilding it from memory takes longer than building it did.

The hour a month after launch

Workflows drift. Vendors update models, an app changes a form field, a new service line brings enquiries the instructions never mention. Someone needs to spend about an hour a month reading a sample of outputs and the error log. If nobody's diary has that hour in it, the workflow has an expiry date. Setting up human review without slowing down shows how to size the sample.

A web design studio's enquiry triage, from start to finish

This is an illustration with made-up but realistic numbers. Say a five-person web design studio (two founders, two designer-developers and a project manager) gets about 40 enquiries a month through its website form. One founder reads each, looks at the prospect's current site, decides whether it's a fit, and replies with questions. It's slow, and good leads sometimes wait two days.

Weeks 1-2: baseline. The founder logs every enquiry. Result: 40 enquiries, about 25 minutes each (roughly 17 hours a month), 6 of them spam or job applications, and a median reply time of 31 hours.

Week 2: tool decision. The studio already runs on Google Workspace. The founders choose Zapier Professional at $29.99 a month billed monthly: a new form entry triggers a Zap, an AI step (AI by Zapier, which needs no separate API key) sorts the enquiry into "good fit", "maybe", "not for us" or "unclear" and drafts a reply, Gmail's Create Draft action saves it in the shared enquiries mailbox, and a row is added to a tracking sheet. Nothing is sent automatically.

Week 3: connect. A company-owned Zapier account, connected only to the shared enquiries mailbox and the form tool, with both founders able to log in.

Week 4: instructions and testing. The founder writes a one-page brief (who the studio works with, typical budgets, services it doesn't offer) and runs it against 30 past enquiries. It sorts 27 correctly. The three misses are instructive: an existing client's support request treated as a new lead, a charity asking for free work labelled "good fit", and a one-line "how much for a website?" labelled "not for us". Fixes: enquiries from existing clients' email domains go straight to the project manager, and anything under 20 words is marked "unclear".

Here's a short extract from that brief and what came back for one test enquiry (illustrative):

We build and maintain websites for small service businesses.
Typical budget: $4,000-$15,000. We don't do online shops over 50
products or logo-only work. Never promise dates or prices; the
founders set those on a call.
Label each enquiry: good fit / maybe / not for us / unclear.
Then draft a reply asking the two or three questions we'd need.

For an enquiry reading "We're a three-chair dental practice wanting a new site with online booking, budget around $8k, ideally live before our rebrand in March", the label was "good fit" and the draft asked sensible questions about the booking system and existing content. Then it ended: "We can certainly have your new site live before March." That's a date nobody had agreed, in plain breach of the brief. The fix was one more line at the top ("The draft must not mention dates, deadlines or availability") and a re-run of all 30 test cases to check nothing else broke.

Weeks 5-6: parallel run. The founder edits each draft before sending and logs editing time. By week 6 it averages about 8 minutes per enquiry.

Handover. The project manager owns the workflow, checks five drafts a month against what was actually sent, and watches the Zapier error alerts, which go to her, not to a general inbox.

At around 11 hours saved a month, the 34 hours of set-up time are recovered in about three months, before counting any extra work won from faster replies.

What you should be holding at handover

Whether you build it yourself or someone builds it for you, don't call the implementation finished until you have all of these:

• A list of every account involved, who owns it, and which card pays for it.
• A one-page map of the process as it now runs, including where a person checks the output. If the process isn't written down at all yet, start with documenting your processes before adding AI.
• The instructions or prompt text, stored in a shared document with the date of the last change.
• The test log from piece 4, so future changes can be re-tested against the same cases.
• A run-book: how to pause the workflow, what to do when it errors, who to call.
• The baseline figure, the current figure, and the date of the next review.

The run-book is the item most often missing. For the studio's triage workflow, it fits on half a page (illustrative):

RUN-BOOK: Website enquiry triage           Owner: project manager
What it does: new form entry -> AI sorts it and drafts a reply ->
  draft saved in the shared enquiries mailbox -> row added to the
  tracking sheet. Nothing is sent automatically.
Accounts: Zapier (company account; both founders and the PM can log
  in; company card). Google Workspace shared enquiries mailbox.
To pause: open the triage Zap in Zapier and switch it off. Enquiries
  still arrive in the mailbox as normal.
If drafts stop appearing: check the Zap's history for errors, check
  the form tool is still sending entries, tell a founder the same day.
Monthly: read five drafts against what was sent; note tasks used
  against the 750 limit; log any enquiry it labelled wrongly.
Prompt last changed: [date], by [name]. Test cases: "triage-tests" sheet.

The test that it's good enough: a founder who has never opened Zapier could pause the workflow from this page alone.

How to tell, 90 days later, that it actually worked

Launch day proves very little. The real test comes three months in, when the novelty has gone and the workflow has met a busy fortnight. Check four things:

1. The number still beats the baseline. Re-measure the same thing you measured in piece 1, the same way. If editing time has crept back up towards the old figure, the instructions have drifted out of step with the work.
2. The checker is still checking. Look at whether the monthly sample actually happened. If it stopped in month two, you have an unreviewed workflow, whatever the run-book says.
3. Nobody is working around it. Ask the people involved whether they ever skip the workflow and do the task the old way. Workarounds are the earliest sign that a category of cases isn't handled.
4. Errors are being seen. Open the error log. Zero errors in three months usually means the alerts are going somewhere nobody looks, not that nothing failed.

If two or more of these fail, go back to piece 4 with the cases that caused trouble, rather than adding a second workflow on top.

Run against the studio's triage workflow at day 90, the check might come out like this. Editing time has crept from 8 minutes to 12, because the studio started offering website care plans in month two and the brief never mentions them, so every care-plan enquiry needs its reply rewritten. The project manager's sample happened in months one and two, but not three, when a launch took over her diary. One of the designers has been answering referrals from past clients straight from his own inbox, since those people never use the form, so a whole category of enquiry sits outside the tracking sheet. The fourth check passes for the right reason: the error log shows two failed runs from the week the form tool renamed its "Budget" field, and the project manager saw both alerts and re-mapped the field the same day. Three flags out of four sends the studio back to piece 4. The fixes are small (add care plans to the brief and re-test, put the monthly sample on a fixed day, forward referrals to the enquiries mailbox), but without the check none of them would have surfaced until a good lead went unanswered.

Doing it yourself or bringing someone in

A capable owner can do all six pieces for a workflow that lives inside one or two tools they already know. Outside help earns its fee when the workflow spans three or more systems, touches client money or regulated advice, or when you've started twice and stalled at the testing stage. What an AI implementation consultant actually does sets out what to expect, and done-for-you versus done-with-you implementation helps you decide how involved your team should be. If you'd rather map the six pieces with someone, that's what my AI implementation consultation covers.

Further reads

• How Long Does AI Implementation Take for a Small Business? — Realistic timelines once you know the pieces of work.
• Hidden Costs of AI Implementation Most Small Businesses Miss — The costs that sit outside the subscription price.
• AI Implementation Roadmap for Small Businesses: 5 Phases in 90 Days — Sequence several workflows over the first 90 days.
• Small Business AI Implementation Checklist: Before, During, After — A before, during and after checklist to tick off.
• How to Pilot AI in Shadow Mode Before Customers See It — How to run the parallel test safely before customers see it.
• How to Evaluate an AI Implementation Proposal or Quote — Judge an outside quote against the work described here.
• Why AI Projects Fail in Small Businesses (It's Rarely the Tech) — The seven organisational reasons small-business AI projects fail, what each looks like by week three, and an eight-question check to run before you start.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Zapier pricing page and help pages on AI by Zapier and Gmail actions; ChatGPT Business and Google Workspace pricing pages (checked September 2026).