The Handoff Integrity Test: When AI Tool Sprawl Turns Your Team Into Middleware

Six · 11 min read · May 7, 2026

The Handoff Integrity Test: When AI Tool Sprawl Turns Your Team Into Middleware

Say your team is getting ready to pilot a new onboarding offer.

One person used an AI research tool to pull findings out of a set of customer interviews. Another had a writing assistant turn those findings into a recommendation: aim the pilot at customers in their first month, and test annual billing. The plan sits in your project tracker. In a chat thread, someone from finance pointed out that none of the interviewees buy through a procurement team, and the billing idea got pulled back.

Every tool did its job. Then the person writing the pilot emails gets a link to the recommendation and a note that says, "Here's the analysis."

Now they have to work out what was decided. Is the segment settled, or still open? Which interviews is the recommendation based on? Who approves the emails? Does anyone know finance pushed back on billing? The answers are spread across four tools and a few people's memories, so the writer messages you.

So you paste the audience definition into another chat, dig up the interview the recommendation leaned on and forward the finance thread. Nobody owns the space between the tools, so it falls to whoever remembers the most.

In software, middleware is the layer that sits between systems and passes information from one to the next. In the pilot, that layer is you.

Sprawl is more than a tool count

If you've been asked to get AI tool sprawl under control, you were probably handed an inventory problem: how many tools, seats, agents and renewals the company has, what they cost and whether anyone can govern them. Okta's Businesses at Work 2025 report, drawn from anonymized data across Okta's own customers, put the global average at 101 apps per customer in 2024, counting every app, not just AI tools. The pilot used four, and the count isn't what sent the writer to you.

Inventory is the easiest kind of sprawl to count. The pilot shows two kinds that are harder to see. The recommendation passed through a research tool, a writing assistant and a tracker before it reached the writer, and someone had to move the information along at each step. Call that workflow sprawl. Then there's decision sprawl: the evidence sits in one tool, finance's objection in another and the approval nowhere in particular, so nobody holds one account of what the team decided.

There's also a cost that comes before the handoff to the writer. DORA, which researches how software teams build and ship, writes in an article on balancing AI tensions that the effort of choosing between and managing "multiple, disconnected internal AI tools" creates "a new layer of decision-making toil." Your team paid it twice, choosing a tool for the interviews and then one for the recommendation, all before the writer saw a link.

Some of this work is the job

Being the person in the middle isn't always waste. Some translation is judgment. Deciding whether finance's objection changes the pilot is real work, and it belongs with someone who understands the business. So does deciding which customers the offer is really for.

The waste starts when capable people keep doing information logistics that the system could have kept for them:

Switching tools isn't the problem on its own. Researchers from Autodesk, Simon Fraser University and the University of Toronto interviewed 15 knowledge workers and five product teams about why people switch applications partway through a single task. People switch deliberately, they found, for reasons that come from the tool, the person, the workflow and the content. It's a small study. The pilot's research tool and writing assistant had good reasons to be there too: one pulled findings from the interviews, and the other turned them into a recommendation. The trouble started when that recommendation reached the writer.

So the useful question about a handoff is whether the work survived it.

The Handoff Integrity Test

A handoff is intact when the person receiving it can carry on responsibly without piecing the earlier conversation back together from scattered files and memory. The test asks five questions:

  1. Decision: is the decision named, or the exact question that's still open?
  2. Evidence: can the receiver trace it back to the source it rests on?
  3. Owner: is one person accountable for what happens next?
  4. Next action: is the next step spelled out?
  5. Exception: are the doubts, objections, limits and stop conditions visible?

Use it before you buy another connector or blame the model. It shows you where people are quietly making up for what the handoff lost.

Decision: did the conclusion survive?

"Here's the analysis" isn't a decision. The writer needs to know whether they're meant to carry the recommendation out, weigh it, revise it or choose between options, and which earlier choices are already settled so they don't reopen them by accident.

For the pilot, an intact handoff would say something like this:

Decision: Aim the pilot at customers in their first month. Pricing is still open. Don't reopen the segment unless new evidence contradicts the interviews we approved.

Evidence: can the receiver trace why?

The writer needs to know which interviews the recommendation rests on, how old they are, what they actually support and where they run thin, so the emails don't promise more than the customers said.

This gets harder when AI writes the step in between. The writing assistant can keep a sentence from the research and lose the interview that made it defensible. NIST's AI Risk Management Framework asks teams to document how an AI system's output may be used and overseen by people, with "sufficient information to assist relevant AI actors when making decisions and taking subsequent actions."

Owner: does one person hold the next decision?

A channel, a department or a group can take part in a decision, but none of them can settle a question by itself. If the writer has to find out again who owns the pilot after every handoff, the work is moving around without anyone taking responsibility for it.

"Marketing owns it" is usually too broad. "The product marketing lead approves the claims, and legal reviews the regulated wording" tells the writer who decides and who gets consulted.

NIST's framework says the same about AI risk work: roles, responsibilities and lines of communication should be documented and "clear to individuals and teams throughout the organization." When ownership is left implied, the writer ends up asking you.

Next action: can the receiver move without a re-brief?

The writer should be able to start drafting without a meeting to work out what you meant. Say what needs doing, to what, within which limits and how they'll know it's done.

Next action: Draft the five-email pilot sequence for the approved segment, using the approved claims. Send it to the product marketing lead for review before anything is scheduled.

DORA's value stream mapping guide warns that handoffs between teams "are often sources of miscommunication and delays." Its guidance on visibility of work suggests tracking how often a step receives something from the step before it that it "can use without requiring rework." In the pilot, that's whether the writer can draft the emails from what arrived or has to message you first.

Exception: did the friction travel too?

The exception should be as easy to see as the recommendation. In the pilot, that's finance's point that none of the interviewees buy through procurement, and the fact that nobody has settled pricing. Written down, it might look like this:

Exception: Two interviews support annual billing, but none of the interviewees buy through procurement. Don't turn this into a pricing recommendation without evidence from finance and sales.

AI tools are good at tidy summaries, and a tidy summary is exactly where finance's objection gets smoothed away. A clean handoff that drops the reason for caution can do more harm than a messy thread that kept it.

Scoring the test

Give one point for each of the five that the receiver can find without reopening the earlier conversation. A 5 doesn't prove the work is right, only that someone can check it.

Score What it tells you What to do
5 The handoff is intact and can be reviewed Carry on, and check the work against its sources
3 or 4 The work can move, but something is still missing Fill the gap before you automate the transfer
1 or 2 The receiver has to rebuild important judgment Redesign the handoff before you add another tool or agent
0 A file moved, but the work didn't Stop and settle the outcome, the evidence and the owner again

The handoff in the opening scores 0. "Here's the analysis" doesn't say what was decided, point to an interview, name who approves the emails, give a next step or mention finance. None of the tools failed, and the writer still had to message you. To score the whole workflow, do the same at every handoff from the first interview to the emails going out.

Audit one workflow from start to finish

Pick a real outcome, not your software list, and choose one where the pain is already showing. For the pilot, that's the path from the interviews to the approved emails. Another team might pick customer feedback on its way to a product decision.

Start with whatever created the work and follow one recent item until someone accepts it, sends it or acts on it.

Then map five things:

  1. What started the work?
  2. Which people, tools, agents and data stores touched it?
  3. Which of the five survived each handoff?
  4. What did someone have to search for, copy, restate, reconcile or ask about again?
  5. Who accepted the final result, and against what standard?

Don't start with login time or license cost. Start with the moments someone had to rebuild the work, like the writer's message asking what was decided. Those moments show you where people are acting as adapters between tools.

Connecting tools isn't the same as keeping the work intact

Connecting the pilot's tools would help less than you'd think. The recommendation could flow from the writing assistant into the tracker automatically and still arrive without the interviews it rests on, the name of whoever approves it or finance's objection.

The same holds as AI agents start handing work to each other. Google's Agent2Agent protocol lets agents advertise what they can do, send each other messages carrying "context, replies, artifacts, or user instructions," and follow a task through its lifecycle with state updates, "even when they don't share memory, tools and context." That can cut down on one-off integration work.

But a protocol can't decide which interview the recommendation should rest on, who owns the pilot or whether finance's objection should hold it up. It can deliver the recommendation to an agent drafting the emails. It can't tell that agent the pricing question is still open unless someone wrote that down. People still make those calls when they design the work.

Should you consolidate your AI tools?

Fewer tools can help, as long as that means fewer places where people have to rebuild the work. Consolidate when duplicate tools, scattered governance or handoffs that keep breaking cost you more than the specialized tools are worth. Keep several when each has a clear job and the handoffs between them hold up. The pilot's research tool and writing assistant each did their job, so retiring one wouldn't have spared the writer that message to you.

For each tool, you have four choices:

What to ask an AI vendor about handoffs

A demo shows you what one agent can produce. Questions about handoffs tell you whether its output can go into a real workflow without someone on your team exporting it, routing it and keeping the paper trail by hand. Ask:

A good answer doesn't have to be "everything lives here." It has to explain how the work survives when it goes somewhere else.

Stop buying tools to fix missing decisions

Run the test on one workflow. Pick the one where someone most often asks "what did we decide?" and score each handoff in it. At every step, ask one question: could the next person carry on without messaging you first?

Say the pilot ran on Supanova atoms. The interview notes go on the project, and the atoms working on it can read them. When the recommendation's billing line needs evidence from finance and sales, you can highlight it and ask for a follow-on task. And when the five emails are drafted, they come back for your review, the same as any finished work.

Someone on your team still decides whether the pricing question stays open. Your team should make the decision. It shouldn't have to be the middleware that keeps the decision alive.

See how it works at supanova.team.

Sources

AI Tool SprawlAI OperationsAI HandoffsAI WorkflowsTeam Coordination