Usually no. Before you buy anything, list the AI features already switched on in your email, accounting, and CRM software. Test each one against a real workflow that has a named owner and a pass or fail check. Spend money only on the one step that still can't be finished inside the tools you already run. Most small businesses skip this inventory. They end up paying for a second tool that does what the first one could already do, and the step that was actually blocking them stays blocked.
The claim: your stack can probably finish more than you think, and less than the vendors say
Here is the position, stated so you can prove it wrong. For a typical small business, the AI built into existing software can fully finish at least one recurring workflow today, with no new purchase. It will also stall at a predictable point: the handoff between two systems. That handoff is the only place where a new tool earns its keep.
You can test this claim in two weeks. If your inventory turns up nothing that finishes a real task, or if the embedded features fail inside a single system and not at the handoffs, then I'm wrong about your business and you should shop. My bet is that you won't get that result.
The reason is structural. Vendors build AI features to keep you inside their product. Microsoft puts Copilot into Outlook so you draft replies there. Intuit puts Intuit Assist into QuickBooks so you categorize transactions and chase invoices there. HubSpot puts Breeze into its CRM so you summarize contacts and write follow-ups there. Each feature is tuned for work that starts and ends in that one product, and each one gets vague when the work needs to leave it.
Step one: make the list
Your first job is an inventory, not an evaluation. Open every tool your team logs into weekly and write down every AI feature it offers, whether or not anyone has turned it on.
Go through admin settings, not marketing pages. Marketing pages describe the roadmap. Admin settings describe what your license actually includes today. For each feature, write down three things:
- **What it claims to do**, in the vendor's words. - **Whether it's on**, and for whom. - **What data it can see.** Just the current email thread? The whole mailbox? One customer record, or the full CRM?
That last column matters more than it looks. An email assistant that only sees the open thread can't tell you that this customer complained last month. A bookkeeping assistant that can't see your CRM doesn't know which invoices belong to your most important account. The data boundary is usually where the feature's usefulness ends.
For a five-person services firm, the list tends to look like this: a drafting and summarizing assistant in Gmail or Outlook, a meeting recap feature in Zoom or Teams, transaction categorization and invoice reminders in QuickBooks or Xero, and contact summaries plus email sequences in the CRM. That's four or five features, all already paid for, and often nobody owns any of them.
Step two: pick real workflows, not demos
Test each feature on a workflow your business already runs, with a named person who owns the result. Demos show a feature at its best. Your Tuesday afternoon shows it as it is.
A workflow for this test has four parts:
1. **A trigger.** Something that already happens: a new inquiry arrives, a month closes, a client meeting ends. 2. **An owner.** One person, by name, who is accountable for the output and who reviews it. 3. **A finished state.** What "done" looks like in plain words. Not "helps with invoicing" but "every overdue invoice older than the due date has had a reminder sent." 4. **A pass or fail check.** A yes or no question the owner can answer in under a minute.
Good candidates are the dull, frequent ones. After every client call, a recap goes into the CRM record and a follow-up email is drafted. At month end, uncategorized transactions get a suggested category and the owner approves or fixes each one. When a new inquiry comes in, the reply is drafted within the hour using your standard answers.
Bad candidates are the ones people get excited about: "use AI for strategy," "have AI analyze our business." Neither has a trigger, a finished state, or a check, so neither can fail. A test that can't fail tells you nothing.
Step three: run each test for two weeks and score it honestly
Give each pairing of feature and workflow two weeks of real use, then score it as one of three outcomes. Two weeks catches enough repetitions to show patterns without dragging on so long that the owner stops paying attention.
The three outcomes:
- **Finishes.** The feature completes the workflow, the owner reviews it quickly, and the check passes almost every time. Keep it and write down the steps so it survives the owner's vacation. - **Assists.** The feature does part of the job, but a person still does meaningful work to reach the finished state. This is the most common result, and it's fine. Record exactly which step the person is still doing. - **Fails.** The output is wrong often enough that checking it takes longer than doing the task by hand. Turn it off for this workflow and move on.
For illustration, suppose a design studio tests the meeting recap feature in Teams on its client calls. The recap is accurate and the drafted follow-up email is usable. But someone still copies the action items by hand into the project tracker and the CRM. Score: assists. The step that didn't get done is "move action items from the meeting into the systems where work is tracked." Write that sentence down. It's the most valuable output of the whole exercise.
Keep the owner's review time as your real measure. If checking the output takes longer than producing it by hand used to, the feature fails, however impressive the draft looks. A rough weekly tally is enough. The goal is a clear before and after for one workflow, not a dashboard.
Where embedded AI stops
Embedded AI almost always stops at the edge of its own product. After a few rounds of testing, you'll see that the "assists" results share a cause: the unfinished step moves information from one system into another.
This is predictable, because each vendor's assistant was built to see its own data. Outlook's assistant knows your mail. QuickBooks' assistant knows your ledger. The CRM assistant knows your contacts and deals. None of them was built to read an email, check it against an open invoice, and update a deal stage in a third product. Some vendors offer connectors that close part of the gap. When they work for your case, use them before you buy anything. But connectors between competing products tend to be shallow, and the most useful context often doesn't cross them.
A few other limits come up often enough to plan for:
- **The feature can't act, only suggest.** It drafts the reminder but can't send it on a schedule without someone clicking. Sometimes that's a setting you haven't found. Sometimes it's a deliberate product choice. - **The feature doesn't know your rules.** It categorizes transactions by general patterns, not by the fact that your two biggest clients get billed under a separate project code. - **The feature changes under you.** Vendors update embedded AI on their own schedule. A workflow that passed in March can drift by June. That's why the pass or fail check stays in place after the test ends, not just during it.
None of these is a reason to buy something new yet. Rules can often go into templates, saved prompts, or custom fields. Suggest-only features can still finish a workflow if the owner's click is the review step you wanted anyway.
The one gap that justifies buying
Bring in a new tool only when you can name a specific step, in one sentence, that none of your existing tools can finish, and the step recurs often enough to matter. Everything else is shopping.
The sentence should read like this: "After a client call, the action items need to land in the project tracker and the CRM deal record without anyone retyping them." That gives you something to test a vendor against, a place to measure the new tool, and a way to tell whether it worked. If the sentence you write is "we need better AI," you haven't found the gap. Go back to the test results.
When you do buy, buy narrowly. The right purchase usually connects systems (an automation platform like Zapier or Make, or an assistant that can read from and write to several of your tools) rather than duplicating what you already have. You don't need a second email drafter. You need the step between the email and the ledger.
Hold the new tool to the same standard as the embedded features: a named owner, a two week test, a pass or fail check on the one step it was bought to finish. If it doesn't finish that step, it goes the same way as any feature that failed.
A 30/60/90 day sequence
The whole process fits in a quarter, and most of the value arrives in the first month.
**Days 1 to 30.** Build the inventory. Pick three workflows, each with a named owner and a written pass or fail check. Run the first two week tests. By day 30 you should have at least one workflow scored as "finishes" and running without new spending.
**Days 31 to 60.** Test the remaining feature and workflow pairs. Write the procedures for anything that finishes. Collect every "assists" result and write down, in one sentence each, the step a person still does. Group those sentences and look for the ones that repeat.
**Days 61 to 90.** If one unfinished step keeps showing up and costs the owner real time each week (say around 2.5 hours, calibrated to your own tally), that's your gap. Evaluate tools against that single step. If nothing repeats, don't buy. Keep running what works and repeat the inventory next quarter, because vendors will have shipped new features by then.
What you end up with
At the end, you have a short list of owned workflows that your current software finishes, a clear record of where it stops, and at most one purchase aimed at a named gap. That's a better position than most businesses reach after buying three AI tools.
It's also easier to defend. When someone on the team asks why you aren't using the latest assistant they read about, you can point to the workflow, the owner, the check, and the result. If the new assistant finishes a step your stack can't, it gets a two week test. If it doesn't, you've already answered the question.
Start with what you're already paying for. Buy only for the step that still won't finish.
Related
- integrate-ai-into-existing-business-processes - smb-ai-strategy-one-workflow - measuring-time-saved-on-one-ai-workflow