The Model Is the Easy Part. The Integration Is the Project.
Five days ago we argued that you should stop looking for an AI project and pick a workflow instead. Say you did that. You picked supplier invoice exceptions, or quote follow ups, or support triage.
Here is the part that surprises people. Almost none of the work left in front of you is AI work.
The failures are architectural, not model failures
In the 2026 State of AI Agents report, integration with existing systems came out as the most cited obstacle for teams building agents, named by 46 percent of respondents. Not hallucination. Not model quality. Plumbing.
Gartner expects more than 40 percent of agentic AI projects to be cancelled by 2027. One compilation of 2026 benchmarks found that roughly a fifth of deployments report negative ROI, and put it plainly: those teams did not lose a model fight, they lost a scoping fight. Deloitte’s 2026 figures point the same way, with 84 percent of companies not having redesigned any actual job around what AI can now do.
Read that as good news. The risk in your first embedded AI workflow is a kind of risk you have handled before. You have integrated systems. You have scoped projects. This is that, with a model somewhere in the middle.
Five questions that are the actual project
If you cannot answer these, you do not have a project yet. You have an intention.
Where does the input come from? Structured data already sitting in your ERP is cheap to work with. A PDF that forty suppliers each format differently is not. This one answer will move your estimate more than any other.
Where does the output land? The most commonly skipped question, and the one that quietly kills adoption. If the AI’s answer shows up in a separate chat window that someone has to remember to visit, you have built a second system, and people will route around it inside a month. The output has to appear where the work already happens: a field, a flag, a draft, a queue, on the screen your team already has open.
Who is allowed to act on it? Suggestion, or action? For a first deployment the answer is suggestion, with a named human approving. Autonomy is earned per task, not switched on at launch.
What happens to the ones it gets wrong? Every flow needs an exception path an ordinary user can follow without special knowledge. If the answer is “we will check the logs”, you do not have one.
Who owns it in six months? Prompts drift, suppliers change their formats, someone swaps the model to save money. Name the owner now, while naming them is still free.
What smaller companies should take from this
Benchmark compilations put median time to value on agent deployments at around five months, and closer to nine for finance and operations work. Plan against that, and understand that most of it is integration time, not model time.
The advantage a 40 person company has is that it can answer all five questions in a single meeting. There is no architecture board. The person who feels the pain, the person who owns the system and the person who signs the cheque are often three people who already eat lunch together. Use that. It is a real structural edge over a company a hundred times your size, and it quietly expires as you grow.
Our own ERP work follows the same shape. The valuable part of putting AI inside Etendo was never the model choice. It was deciding where output lands, which role can see what, and what a human does with a low confidence result. We have written separately about why AI projects fail on delivery rather than technology, and this is that argument one level down, in the plumbing.
So pick the workflow, then budget for the pipes. If you want a second opinion on which flow is cheapest to embed into the systems you already run, that is exactly what our free assessment is for.