Why AI Projects Will Increasingly Look Like ERP Projects

The technology is the easy part. AI projects are converging on the shape of enterprise software implementations, which is an uncomfortable comparison and a useful one. This piece draws it out, using FISTA Solutions' AI enablement delivery experience.

Where does the effort actually go?

The distribution that surprises teams.

Activity	Share of effort
Process understanding and redesign	Large
Integration with systems of record	Large
Data remediation	Large
Change management and training	Moderate
Evaluation and quality assurance	Moderate
Building the AI capability itself	Small

Why does the comparison hold?

Because both are business change programmes wearing a technology label.

An enterprise system implementation succeeds or fails on whether the organisation changes how it works. The software is a prerequisite, not the substance. The same is true of any AI deployment that genuinely alters a process.

That is why the failure patterns rhyme: insufficient sponsorship, scope expanding beyond what the organisation can absorb, data problems discovered late, and users reverting to the old process. See the second wave of AI adoption.

What does process redesign involve?

Understanding how the work actually happens, which is never how it is documented.

Every process has undocumented exceptions, informal workarounds, and rules held by experienced people. Automating the documented process produces a system that handles the easy cases and fails on everything else.

Mapping the real process requires time with the people who do it, and it is the step most often compressed. Projects that skip it discover the exceptions in production. See forward deployed engineer knowledge transfer.

Why is integration so large?

Because value requires reading from and writing to the systems of record.

A system that produces a good answer which someone then types into another application has automated the thinking and left the work. The value comes from closing that loop, which means integration with systems that are frequently old and poorly documented.

This is the most reliably underestimated part of any such project. See AI integration with legacy systems.

What about data?

It is worse than the organisation believes, consistently.

Duplicates, contradictions, stale records, and missing metadata all surface when a system starts reading the data systematically. Humans compensated for these silently; software does not.

Budget for remediation as a workstream, not a task. And measure the corpus before committing to a timeline. See why data quality decides AI outcomes.

What organisational dynamics appear?

The predictable ones, in the predictable order.

Enthusiasm during selection, resistance when the process change becomes concrete, a difficult period after go-live, and either stabilisation or reversion depending on whether the change was supported.

The people whose work changes need to be involved early and told the truth about what happens to their roles. Projects that handle this honestly do better than those that manage communications. See AI adoption strategy.

What is genuinely different?

Two things, both favourable.

AI systems can be deployed incrementally in a way monolithic enterprise systems could not. One workflow at a time, with the old process still available, is a far safer path than a cutover.

And they behave probabilistically, which means acceptance testing is insufficient and continuous evaluation is required. That is more work, and it also means quality problems are detectable rather than latent. See how to monitor AI quality in production.

What is the counter-argument?

The counter is that the comparison is defeatist and discourages ambition, since enterprise implementations have a poor reputation. The intent is the opposite: those projects fail from predictable causes, all of which are addressable, and knowing the shape of the work is what makes it manageable.

What does this change for engineering teams?

It means engineering is a minority of the project and should be planned as such. Integration and data engineering dominate, and the AI-specific work is comparatively small.

It also means engineers need to work alongside process owners rather than receiving requirements, because the requirements do not exist until the process is understood.

What does this change for buyers?

It means evaluating vendors on integration capability and change support rather than on model quality.

And budgeting your own people's time realistically. A vendor cannot redesign your process or clean your data, and projects that assume otherwise stall.

What should leaders do about it now?

Sponsor it as a business change programme with an operational owner, not as a technology project. That single framing decision predicts the outcome better than any other.

Then phase it. One workflow, delivered and stabilised, before the next.

Does this apply to agents?

Yes, with additional weight on permissions and exception handling. An agent acting within a process needs the process's approval chains and escalation paths encoded, which requires the same process understanding.

The incremental deployment advantage is strongest here: an agent handling ten percent of cases, with the rest going to people, is a low-risk starting point. See AI pilot checklist.

How will you know if this is happening?

Watch for timelines set from the technology estimate, for data problems discovered in month three, and for the project owned by a technology function. Each predicts the familiar difficulties.

How FISTA Solutions reads this

FISTA Solutions builds and operates production AI systems through AI agents, AI enablement, and forward deployed engineering: process understood before it is automated, integration and data remediation scoped as the majority of the work, and delivery phased one workflow at a time, decisions documented with their reasoning, and handover that leaves your team able to maintain what was delivered. The record is 150+ projects for 50+ companies across 12+ countries.

To discuss what this means for your roadmap, message FISTA on WhatsApp, or read AI pilot to production.