🧭 Faster, new, or different?


Sep 20, 2026 | Read time: 11 mins.

Kevin Donovan · Architects' Hub

Hey Reader and welcome to the Hub.

This week I’ve been working through a distinction that can get lost in an AI progress report.

People can use more AI without changing how the organization works. They can also change how they work in ways the report barely notices.

Both can happen in the same company, at the same time.

The useful question is what changed, what evidence supports it, and whether another team could repeat it.

In this edition

  • The 3 kinds of change hiding under “AI maturity.”
  • The tell, the convincing deepfake, and the check for each.
  • A worked example: faster service, a new offer, a different workflow.
  • A reader’s Christensen question: build inside the company, or give the work a separate home?
  • A 30-minute check before asking another team to copy it.
  • Community Corner: which condition is missing where you work?

Let’s get into it.👇

________

🧭

________

​

Same dashboard. Different work.

Leadership asks whether AI is making a difference. The question lands with IT.

IT shows licenses bought, people onboarded, weekly usage, and token consumption. Those numbers answer sensible questions about deployment and adoption.

They leave another question open: what can the organization now do?

Imagine 2 teams, both reporting 80% AI adoption.

In the first, people draft, analyze, and write code faster. Work still passes between the same people, through the same approvals.

In the second, people direct a shared workflow. Routine steps run together. People review results and handle exceptions that need judgment.

The dashboard can make them look much the same:

A useful opening question is: if AI access disappeared tomorrow, what would break?

“We’d do the same work, but slower” points to individual productivity. That’s a real gain.

“We’d have to rebuild how we deliver” deserves a closer look.

Dependence alone doesn’t prove transformation. A team can become dependent on a tool while keeping every handoff and approval it had before.

Follow a piece of work from request to result. Look at who acts, who decides, and where it waits. Then ask whether the changes improved delivery.

That’s where the evidence starts.

________

🧭

________

​

3 levels, without the automatic promotion

I’ve been calling these levels, but they’re different kinds of change. A company can have all 3. A higher number isn’t automatically the right goal for every workflow.

1️⃣ Individual: the same work, faster

An architect gets a useful first analysis sooner. A developer spends less time writing routine code. A team prepares the same report in fewer hours.

Measure the gain, including review and rework. Time saved on a first draft can disappear if someone else spends longer correcting it.

2️⃣ Business: something new becomes viable

AI makes a product, service, offer, or price point possible that previously didn’t make sense.

The offer doesn’t have to be entirely new. Perhaps you can now serve smaller customers because the cost of delivery has changed.

The evidence belongs with the offer: who uses it, what they receive, and what makes it viable.

3️⃣ Operating model: the way work gets done changes

Handoffs disappear. Responsibilities change. Decision rights become explicit. People direct work and handle exceptions instead of moving every case through the same sequence.

A different workflow still has to earn its place. Fewer steps are useful only if the result holds up.

10,000 people saving an hour a week is a substantial productivity gain.

Turning some of that capacity into a new service takes a business decision. Changing how the service runs takes design work, ownership, and changes across boundaries.

Scale the evidence you have. Don’t promote the claim just because the numbers grew.

The tell and the convincing deepfake

I’m using “deepfake” here as a metaphor: a convincing appearance of progress that the evidence doesn’t support. It doesn’t require anyone to be dishonest.

Level 1️⃣

🫆 Tell: the same work takes less effort or time, with acceptable quality.

🪞 Deepfake: more seats and more tokens. Neither tells you whether the work improved.

✅ The honest check: compare a recurring task before and after, including review time. Ask what the team would do without AI.

Level 2️⃣

🫆 Tell: people use an offer that AI helped make viable.

🪞 Deepfake: a convincing demo presented as a business capability before anyone relies on it.

✅ The honest check: name the offer, its users, and what changed in delivery cost or capability. Check the ongoing effort, too.

Level 3️⃣

🫆 Tell: responsibilities, handoffs, or decision rights changed in the working process.

🪞 Deepfake: a model inserted into one step while every original handoff remains.

✅ The honest check: draw the workflow before and after. Add who decides and who handles exceptions. Timings alone won’t show this.

Activity is easy to count. Consequence takes evidence.

One service, 3 possible changes

Here’s an illustrative example, rather than a client case.

A service team prepares monthly equipment-maintenance recommendations. Someone gathers records, an analyst reviews them, and a service manager approves the recommendation.

AI helps the analyst prepare a first draft. The manager still reviews every report. The same customers get the same service.

That is a candidate for Level 1. Check total effort and quality before claiming the gain.

Now suppose the lower delivery cost makes a smaller monthly package viable for customers the company couldn’t previously afford to serve.

That is a candidate for Level 2. The evidence would include actual customers, delivery costs, and a service they find useful. The cheaper draft alone doesn’t establish it.

Finally, suppose the team redesigns delivery. An approved workflow gathers the records and prepares routine recommendations. Defined exceptions go to the service manager for a decision.

That is a candidate for Level 3. The routing and decision responsibilities changed. You would still check error rates, exceptions, and whether the service improved.

These changes can coexist. They also need different evidence and different owners.

The analyst can demonstrate time saved. A service owner must decide whether the new offer works. Someone with authority over delivery must own the redesigned process.

Calling all of it “AI adoption” makes those decisions harder to see.

________

🧭

________

​

Inside the company, or in a separate business?

​Kevin Broekhoven raised a useful question in the LinkedIn discussion. He connected Levels 2 and 3 to Clayton Christensen’s work: could a separate business protect the new model from resistance inside an established, high-margin company?

That adds an important question to this week’s model: where can the change get a fair chance?

Christensen’s argument goes beyond reluctant people. An established company allocates money and attention through processes built for its existing customers and economics. Those processes can reject a promising business with different needs. An autonomous unit with separate funding is one possible response; a separate legal company is not the only form of separation. Harvard’s account of Christensen’s argument explains that distinction.

Here’s how I’d apply it to our 3 levels.

At Level 2, the case for separation gets stronger when the new offer conflicts with the core business’s economics or incentives.

Take our smaller maintenance package. If salespeople are rewarded for large contracts, and every proposal must meet the premium service’s margin target, the new offer may never get tested fairly. Protected funding, different targets, and authority over delivery could give it room to prove itself.

That doesn’t make every AI-enabled offer disruptive. Christensen’s theory concerns a particular market process, including entry through lower-end or previously unserved markets. The Christensen Institute’s explanation is useful on that boundary.

At Level 3, a protected team can also develop and test a different operating model. But its success alone doesn’t show that the parent company has changed.

If changing the existing company is the goal, name the receiving owner and the changes needed in its funding, access, responsibilities, and decisions. If building a separate business is the goal, continued independence may be appropriate.

So I’d decide what needs protection, what needs to connect, and where the change is meant to last. Reintegration should follow that purpose, rather than be assumed from the start.

Thank you, Kevin, for adding this question. It makes the next step more precise: before asking how the work spreads, decide where it belongs.

________

🧭

________

​

Why the next team can’t simply copy it

A successful team may describe its work as a side experiment or a grassroots effort.

That can describe its standing inside the company more than the work itself. It may have no sponsor or route for sharing what it has built.

Training can help another team learn the tools. It can’t, on its own, give that team access to a system or authority to change an approval.

Before copying the playbook, inspect 4 conditions.

📁 Knowledge a model can read and use

Can the workflow reach the current guidance, definitions, examples, and exceptions? Can people tell which source is authoritative?

Without that, the experienced team fills gaps from memory. The next team copies the workflow and discovers how much was never written down.

Last week’s customer-definition problem belongs here. A model can read “active customer” perfectly and still apply the wrong business meaning.

Start with the knowledge required for one workflow. Settle the terms that affect its decisions, and name who keeps them current.

🔌 Systems the workflow can reach

Can the next team access the required systems through an approved route, with the permissions it actually needs?

The first team may depend on manual exports, personal access, or somebody who knows whom to call. Those arrangements can explain a successful demonstration without providing a repeatable service.

Check where the work reads, writes, and waits. Include what happens when access or an integration fails.

🛡️ Guardrails people can act within

What may proceed? What requires review? What must stop?

“Use your judgment” can work between colleagues who share years of context. It travels poorly to a team that doesn’t know which exceptions matter.

Write the boundary for this workflow. Include where uncertain cases go and how someone can stop or reverse an action.

👤 An owner for the decision

Who can accept the result, change the boundary, or decide what happens after an exception?

A person watching a queue isn’t necessarily authorized to make those decisions.

Name an owner who has the authority, time, and support to act. Check what happens when they’re unavailable.

These 4 conditions are a starting check, not a guarantee. Skills, incentives, economics, and the nature of the work still matter.

They do give you something more useful to investigate than whether the next team is “resistant to AI.”

________

🧭

________

​

A 30-minute check for Monday

Use this check when you want the work to spread inside the existing organization. For a deliberately separate business, first agree which dependencies it should share with the parent.

Bring someone who does the work, someone from the team expected to repeat it, and someone who can change a relevant constraint.

Choose one recurring workflow. Keep the discussion small enough to follow a real case from start to finish.

Minutes 0–5: name the change

Complete this sentence:

For this workflow, AI changed __. We can show that through __.

Classify the evidence as faster individual work, a newly viable offer, or a changed operating model. More than one answer is allowed.

If you have only usage numbers, write “outcome not yet checked.” That’s a useful finding.

Minutes 5–15: follow one completed case

Use an actual request and result. Draw the steps, waits, reviews, and exceptions.

Ask the experienced team where it relies on unwritten knowledge, special access, or a decision someone makes informally.

Then let the receiving team mark where it would get stuck.

Minutes 15–25: check the conditions

For each condition, record ready, missing, or unknown for the receiving team. An unknown needs evidence before it becomes a green box.

Copy these prompts into your notes:

  • 📁 Knowledge: which source would we use, and who maintains it?
  • 🔌 Systems: what access works today, and what still depends on a favor or manual step?
  • 🛡️ Guardrails: which cases can proceed, and which need review?
  • 👤 Owner: who can decide on an exception, and have they accepted that responsibility?

Minutes 25–30: choose one next move

Pick one missing condition that prevents a small, useful trial. Name who can change it and agree when you’ll check again.

Write the next step in this form:

Before the next team runs __, we need __. __ owns that change. We’ll check it on __.

In the maintenance example, perhaps the receiving team cannot read the approved service history. Another prompting workshop won’t fix that.

The next move might be to agree and test read-only access for the defined workflow. After that, try a small set of cases under review.

Agree what you will observe: delivery time, correction effort, exceptions, and the quality of the result. Choose the measures that fit the claim.

Also agree what would make you stop or revise the trial.

The meeting should end with one owned change and a way to check whether it helped.

That’s how a team’s success starts becoming something another team can repeat.

________

🧭

________

​

Community Corner

🧭 Community Corner

Which of the 4 conditions is hardest to get right where you work?

You could reply with knowledge, systems, guardrails, or owner. If you have a minute, tell me where the next team gets stuck.

No company names needed. I’m interested in the condition, and what you’ve tried to change it.

If you use the 30-minute check, I’d also like to hear whether it led to a different next step than you expected.

And the small craft cohort I mentioned last week is still taking shape: working architects bringing live decisions and giving each other’s work proper attention.

If you’d like to help shape it, reply Craft and tell me what kind of feedback you miss in your day-to-day work. No commitment, price, or date attached.

Let me know what you're thinking. I read every reply.

Thank you for being part of Architects’ Hub.

Kevin

________

________

Kevin Donovan​
I write Architects' Hub every Sunday for senior architects and technology leaders.
You're getting this because you subscribed at architects-hub.kit.com.
Know someone who’d find this useful? Forward it to them.

​

Korsvejen 9, Viborg, Viborg 8800
​Unsubscribe · Preferences​

Architects' Hub

A Sunday newsletter for senior and enterprise architects. One decision, one practical model, one tool you can use Monday. Written by Kevin Donovan. Join 2,500+ architects. One email every Sunday.

Read more from Architects' Hub
Three doors slightly ajar in a row. Through them, a telephone, a stack of documents with a magnifying glass, and an arrow curving back on itself. A single coin rests on the floor in front of the doors.

Nov. 22 2026 | Read time: 11 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. The room went quiet Hey , Two Sundays ago, our hypothetical pilot case got its commitment record. Last Sunday it got a walkthrough, and the operating owner signed with 3 tests on record. The AI assistant now drafts delayed-order replies for 3 teams, and the drafts are good. This week someone proposed letting it do more than just draft… In this edition The...

Five simple figures in a ring around a small central box, each holding one object: a magnifying glass, a signpost, a wrench, a mop, a headset. One figure carries a leaning stack of boxes.

Nov 15, 2026 | Read time: 11 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. Tuesday, 10:40 am. Hey ,Last Sunday we left a hypothetical customer-service assistant with a commitment record and unanswered questions about capacity, data and support. This week, move the story forward 3 weeks. The team has agreed to a bounded trial on 1 queue. Agents still review every draft. Wider rollout is waiting. I want to spend this edition with the...

A working terminal with four loose connections labeled Funding, Operations, Data and Authority.

November 8, 2026 | Read time: 9 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. Hey ,In this edition of Architects' Hub: What the service lead heard when the sponsor said, “We should roll this out.” A commitment record filled in for one example customer-service queue, PLUS a blank copy for yours. How to take a conditional decision back to the sponsor who wants a date. Community Corner: Santiago Jesus Lasso Valenciano on the work that...