Why small-business AI projects stall after the demo
A demo proves a tool can do the task. Four things decide whether it ever runs the task: an owner, a baseline, an exception path and access.
The demo goes well. The tool reads a messy email, pulls out the details and drafts a sensible reply. Everyone in the room nods. Two months later, nobody uses it, and nobody can quite say why.
This pattern is common enough that it is worth looking at directly. The usual explanation is that the technology was not ready. More often the technology worked fine and the business was not ready to receive it. The same four gaps come up again and again, and each one can be closed before a project starts.
What a demo proves and what it skips
A demo shows that a tool can do a task once, on a clean example, with someone who knows how to run it. That is a real result. It is also the easy part.
A demo does not show:
- What happens when the input is wrong, missing or strange.
- Who checks the output, and where it goes next.
- Whether the tool can reach the systems it needs without someone pasting data in.
- Whether anyone can tell if it is doing better than the old way.
- Who is responsible when it is slow or wrong on a Tuesday afternoon.
Those five gaps map onto four things a project needs before it starts: an owner, a baseline, an exception path and access.
1. An owner
Every workflow that matters already has someone who cares if it breaks, whether or not their title says so. An automation project needs one named person on your side who is accountable for it. Not a committee, and not "the team."
The owner does four jobs:
- Decides what "working" means for the workflow.
- Tests the system with realistic cases before it goes live.
- Signs it off, or sends it back.
- Notices and acts when it starts to drift.
If nobody has those jobs, the system gets built, nobody trusts it, and people quietly go back to the old way. Nothing visibly fails. It just does not get used.
Test: can you write one name next to each workflow you plan to improve? If you write two names, ask which one would get the call if it broke. If you write "everyone," the answer is nobody.
The owner also needs time. A rule of thumb for planning: the owner should expect to give a few hours in the weeks of testing, and a short check-in each week after launch. If your best candidate is booked solid, the project waits, and it is better to know that on day one.
2. A baseline
Without a baseline, "it helped" is an opinion. With one, it is a number.
A baseline is how long the workflow takes today, how often it runs and where the time goes. It comes from watching real work and counting records, not from memory. Without it:
- Nobody can say whether the tool saved anything.
- The people who liked the old way point to every error.
- The people who championed the project point to every success.
- Both sides are right about what they saw, and the project gets judged on mood.
Here is a made-up illustration of the arithmetic. A business sends 60 invoices a month, and each one takes 20 minutes to assemble and send by hand. That is 1,200 minutes, or 20 hours a month. If the new process takes 8 minutes an invoice, the monthly time is 480 minutes, or 8 hours. The saving is 12 hours, a 60% reduction. Those numbers are invented to show the method. The point is that you can only write that sentence if you wrote the first line before the build.
The baseline should cover a small number of workflows, not the whole company. Three to five is plenty. Each needs a start, a finish, a volume, a time per run and a measurement method written down.
3. A defined exception path
Every workflow has a typical case and a pile of odd ones. The typical case is what demos are made of. The odd ones are where projects stall.
An exception is anything the system should not handle on its own: a missing field, an unusual request, an amount over a limit, a customer who is upset, a document it cannot read. The question is not whether exceptions will occur. It is what happens to them.
An exception path answers four questions:
- How is it detected? A rule (an invoice over a set amount), a confidence signal (the system is unsure), or a person flagging it.
- Where does it go? A named person or queue, not a general inbox.
- What does that person see? The original input, what the system did or proposed, and why it stopped.
- How fast must it be handled? A same-day exception and a next-week exception need different routing.
Without this, one of two things happens. The system handles exceptions badly and someone gets burned, which ends the project. Or it hands everything unusual back with no explanation, and the team ends up doing the work twice. Both look like "the AI doesn't work."
Define the exception path by looking at the last 20 or 30 real cases of the workflow and sorting them. How many were typical? How many were odd, and in what way? That sort is also the best estimate of how much a system can take over.
For a simple view, build a table like this and fill it in for each workflow:
| Situation | Detected by | Goes to | Person sees | Handled within |
|---|---|---|---|---|
| Quote request missing a measurement | Required field empty | Estimator | Original request, draft, missing item | Same day |
| Invoice over an agreed amount | Amount rule | Owner | Invoice and source job record | Before sending |
| Customer reply that sounds upset | Reviewer flag | Account manager | Full thread | Within 2 hours |
| Document the system cannot read | Low confidence | Admin | The document and what was extracted | Next business day |
The details are examples. The structure is the point.
4. Access
A tool that cannot reach your systems cannot do the work. It can only advise. A lot of stalled projects are really access problems.
Typical gaps:
- The CRM is owned by someone who is on leave and cannot grant a connection.
- The accounting system allows only one kind of login, and it belongs to the bookkeeper.
- Data lives in spreadsheets on a laptop.
- The software has no way to connect to other tools, or the plan you pay for does not include it.
- IT or the vendor needs to approve a change, and the queue is long.
Settle these in the first week. Ask:
- Which systems does each workflow touch, and who administers each one?
- Can we get read-only access first? (This is the safest way to begin.)
- Does the software offer scoped accounts, so the tool sees only what it needs?
- Is the data we need in a system, or in someone's head?
- Who approves changes to each tool, and how long does that usually take?
Good practice on access is to start with exports, demos and read-only views, and ask for more only when an approved build needs it. Keep credentials in a secrets manager, never in email or chat, and write down who granted what and when so it can be removed afterward.
A readiness check you can run in a week
Before you hire anyone or buy anything, score each workflow you are considering:
| Question | Yes | Not yet |
|---|---|---|
| Is there one named owner with time in the next 6 weeks? | ||
| Do we have a measured baseline (volume, time, source)? | ||
| Have we sorted the last 20 to 30 cases into typical and exception? | ||
| Is there a named person or queue for each kind of exception? | ||
| Do we know every system involved and who administers it? | ||
| Can we get read-only access to each within a week? |
Every "not yet" is the next piece of work. Closing the gaps is a good use of two weeks and costs far less than a build that stalls. Each gap has a fix:
- No owner. Pick the person who does the work most days and give them the title for this workflow. If they cannot, delay it.
- No baseline. Spend a few days watching and counting. It is the fastest of the four to close.
- No exception path. Sort recent cases and write the table above. Nothing needs to be built to do this.
- No access. Start the approvals now, in parallel with everything else, since they are slow.
The demo is still useful
None of this means demos are a waste. A good demo tells you the task is possible and gives people something to react to. It answers "can it do this?" The next questions are different: who owns it, how will we measure it, what happens when it fails, and what does it need to touch.