ChatGPT
Perplexity
Your first AI deployment works, so the obvious second move is another one in the same department. The data is already connected, and the team knows the tools. That second deployment will pay less than the first. AI returns show up when work moves between departments, and a second deployment inside the same one moves nothing.
An IDC study commissioned by Microsoft found that the companies it classified as leaders were using AI across seven business areas on average, and self-reported a 2.84x return against 0.84x for the slowest adopters. That survey only covered companies with 1,000 or more employees, most above 5,000, so read it as direction and not a target.
The finding worth keeping is simpler. Value appears where one department’s work reaches another, and that connection is also the most fragile thing you will build.
Step 1. Confirm your first deployment is genuinely finished
What this step involves.
Check whether someone outside the build team owns a number the first deployment moves, and whether they looked at that number this month. If the answer is no, you have a successful pilot and not a foundation.
Why teams skip it.
A pilot that works feels complete. Grant Thornton surveyed 950 business leaders and found mid-market companies more than twice as likely as larger ones to be still piloting, while companies running AI fully in production were close to 4 times more likely to report revenue growth.
How to do it.
Name the business owner, not the technical one. Agree on the single metric they are accountable for, then give it one full reporting cycle in production before you plan anything downstream.
How you know it worked.
Someone in the business can tell you what the deployment did last month without asking engineering.
Step 2. Settle the shared definitions in one place, not between two systems
What this step involves.
Sales and finance both use the word customer, and they often mean different things by it. One counts a signed contract, the other counts a paying account.
The fix is one definition held in one governed place, whether that is a semantic layer, a master data setup, or a metadata catalog, so the third and fourth department inherit it.
Why teams skip it.
Reconciling two systems is a week of work. Standing up a place where definitions live is a quarter, and it produces nothing you can demo. So teams do the week, then do it again for the next pair.
Grant Thornton found that 70% of mid-market companies say at least half their core applications are not ready for AI, against 39% at larger companies.
IBM’s research puts a number on what that costs, finding over a quarter of organizations lose more than $5 million a year to poor data quality.
How to do it.
Start with only the entities your first two departments exchange, and build the layer around those. Name one owner per definition. Point both systems at it and stop letting each hold a copy.
How you know it worked.
The second connection takes less definition work than the first. If it takes the same amount of work, you built an agreement and not a layer.
Stay updated with Simform’s weekly insights.
Step 3. Build one place where output gets checked, and route everything through it
What this step involves.
Something has to verify the first deployment’s output before the second acts on it. Build that check once, as a service both deployments call, and send later connections through the same one. A separate validator per connection gives you no view of what is failing anywhere.
Why teams skip it.
The connection point looks like plumbing, and plumbing gets built where it is needed. The clearest evidence on what that costs comes from preprint research on AI agents handing work to each other, so read the multiples as direction.
Across 180 configurations, agents connected without a central check multiplied errors by 17.2 times against a single agent working alone, while routing through one coordinating layer held that to 4.4 times.
A second preprint studying 1,600 execution traces across seven frameworks found failure rates between 41% and 87%, with information lost at connection points as the second most common cause.
How to do it.
Before you build, establish whether the two tasks can run in parallel or whether the second has to reason through the first one’s output. The same research found central coordination helps in the first case and degrades performance by 39% to 70% in the second, where a human at the handover does better. Then log every check in one place.
How you know it worked.
You can see what has been rejected across every connection without asking the team that built it.
Step 4. Track one number that covers both
What this step involves.
Each deployment arrives with its own dashboard. Two green dashboards can sit on either side of a connection carrying nothing. Pick a number that only improves if both work.
Why teams skip it.
Every business case was written per deployment, so every metric is per deployment. Forrester argues that AI’s return problem is really a measurement problem, because companies keep judging AI with the isolated metrics they built for automation work.
How to do it.
Choose something that spans the two departments, like order-to-cash cycle time or time from complaint to fix. Baseline it before the second deployment goes live, since you cannot reconstruct that later.
How you know it worked.
The joint number moves in a way neither individual dashboard explains.
A connected order and inventory workflow cut fulfillment lead time by 70%
A global semiconductor manufacturer ran order routing on spreadsheets and manual validation while inventory sat in a separate set of tools. Sales and fulfillment exchanged data by hand, and every handoff carried its own error risk across thousands of daily orders.
Simform connected the two into one platform with real-time inventory visibility across 200-plus fulfillment partners and 5,000-plus SKUs. Order lead time fell by 70%.
No AI was involved in that work, which is the point. Order lead time is a number neither department owns alone, and it only moved once the connection itself was designed.
When a same-department second deployment is the better move
Two conditions change the answer. If your definitions are far enough apart that reconciling them is a quarter of work and not a sprint, the reconciliation costs more than the connection returns. And if nobody can own the checking, an unattended connection is worse than two separate deployments.
Neither condition is permanent. Both tell you the connection is the third thing you build.
Set a deadline for the connected pair before you expand again
The pressure to expand arrives before the second deployment has proven anything. The same IDC data shows returns flatten as deployments mature, with payback averaging around 15 months, and that number is more useful as a deadline than as an expectation.
Decide now what the connected pair has to deliver inside that window. If it falls short, the work is fixing the connection you have, because a third department inherits every definition problem the second one exposed.
Simform’s enterprise AI adoption assessment covers the readiness work in steps one and two.
