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.
Share
🧭 AI agents and authority: write the action boundary
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 refund proposal, and why the quiet in the room was the right response
The entire Agentic Boundary, worked end to end: Allowed, Review, Stop, Escalation, Recovery, Log
How the Boundary maps to the 5 guardrail fields (from back in September, remember? :-)
A blank Boundary for 1 action your org is debating
Community Corner: Sunil Mishra on recovery, and a question for you
Work with me
Let’s get into it.👇
________
🧭
________
“Why not let it issue the credit?”
Month-2 review. The product manager has been watching the data and says something reasonable:
“It already knows when an order is late. 40% of those customers get a goodwill credit anyway, once a customer service rep reads the case. Why not let it issue the credit itself?”
The math was on his side ...and the room went quiet.
The finance partner was doing mental math around exposure she’d never agreed to carry.
The team lead was picturing wrong credits going out 300 times before anyone noticed.
The data owner was wondering what exactly “knows the order is late” means when the record is 6 days stale.
The architect heard the words “let it issue credit” and started a mental list: a write to the payments system, under whose credentials, logged where, reversed how.
Nobody said no. Nobody could say yes. The request had changed category and character.
For 2 months a human read every sentence the assistant wrote and chose to send it. A wrong draft could be corrected before it reached the customer.
“Let it issue the credit” removes approval. Now the team needs an agreed limit on the money the assistant may move, and a way to put mistakes right after the action has happened.
The feature is 1 API call.
The authority is a new commitment, with an accountable owner.
I’ve heard that silence before AI, especially every time an automation gets wired to move money, delete records or email customers.
It’s when the question changes from “does it work?” to “what is it allowed to do?”
Draft, Recommend, Execute
For each action in a workflow, an AI assistant can hold 1 of 3 levels of authority. The model may be the same. The authority changes.
1. Draft.
It produces a reply that a customer service rep reviews and sends. The rep chooses what reaches the customer. A wrong draft can be corrected before sending.
2. Recommend.
It proposes an action with evidence: “Late by 6 days, no prior credit, recommend $15. Approve?”
The rep still decides. As trust grows, approval can become a reflex, so the evidence needs to stay visible and usable.
3. Execute.
It acts without approval on each case, within an agreed boundary. The owners authorize that boundary and carry the operational consequences.
A wrong action needs a process: detect it, reverse it where possible, and put it right where reversal is impossible.
Our assistant has been at level 1 for 2 months. “Let it issue refunds” is level 3. Proposals can skip from 1 to 3 because the feature looks the same from outside: the assistant already “knows” the order is late, so the credit is 1 more line.
The levels tie back to week 1’s 4 owners. Draft needed an operating owner. Execute needs a decision owner who says what it may do, a resource owner who carries the money it moves, and a data owner who vouches for the record it acts on.
Same 4 names, heavier signature.
The level decides the design. Drafting a reply needs order data. Issuing a credit also needs a payment write, with different failure modes. Whose permissions authorize the read? Whose identity authorizes the write?
A conversation log can show the draft; execution also needs an action audit trail and a recovery path. Agree the authority before designing the write path.
The 3 questions before it acts
Three doors before the coin can pass: who gets the call, what evidence, how to put it right.
These questions need answers from the people who carry the consequences. If the room cannot answer them, the assistant can not have authority to issue credits on its own.
Who carries the consequence when it’s wrong?
Not “who’s accountable” on a slide. Who gets the call?
For the credit: the finance partner carries the money, the team lead carries the customer who got a credit and then a confused follow-up, the data owner carries the stale record that said “late” when the parcel was on the porch.
All 3 need a say before the proposal goes ahead.
What evidence does the action need?
A reply needs reliable evidence. Moving money needs agreed checks before it happens. I
s the order late by the carrier’s record or by the order system’s stale copy? Has this customer already had a credit on this order? Is the amount inside an agreed limit? Is the underlying record current, not just the system’s response?
Each check needs an owner. Last week’s “assumed” column, with money riding on it.
How do we put it right?
A wrong draft can be edited before sending. A wrong recommendation can be declined.
Once a credit has gone out, can finance reverse it? Does someone need to contact the customer? If reversal is impossible, who agrees on how to make it right? If recovery is slow, tighten the allowed case.
The architect helps those owners agree on the answers and turns them into a design: who authorizes the action, what evidence it needs, and what happens when it goes wrong.
________
🧭
________
The agentic boundary, worked end to end
Allowed | Review | Stop Written once, read on a Tuesday, signed by the people who carry it.
Here’s the boundary for a goodwill credit on delayed orders in our hypothetical pilot. On LinkedIn this week, I called it the Agentic Boundary: what the AI assistant may do on its own, and what happens at the edges. Every figure below is illustrative, for the owners to agree.
Stop overrides review. Review overrides allowed.
Missing evidence never counts as a passed check.
Action
Issue a goodwill credit to the customer’s original payment method on an order the assistant has identified as late.
Allowed: the system acts; no human in the step
All of these must hold:
The carrier’s own record, not the order system’s copy, shows the parcel 3 or more days past its promised date.
The carrier record was updated within the last 10 minutes. A fresh system response containing an old record does not pass.
No prior credit or refund exists on this order.
The amount follows the agreed goodwill-credit policy and is no more than $15.
The required account-history check shows no past dispute.
It is a single-shipment order, with no review or stop condition present.
Review: the system recommends; a named human decides
Any 1 of these moves the case to review:
Amount above $15, or the customer asks for a different amount.
Missing evidence, carrier record older than 10 minutes, or the carrier and order records disagree.
Prior credit or refund on this order.
A past dispute on the account.
The order contains more than 1 shipment.
The recommendation shows carrier status and timestamp, order details, prior-credit and dispute checks, the proposed amount, and anything missing. The rep decides after checking that evidence. Record the decision and its reason. A missing check must be resolved before approval.
Stop: no credit is issued, and the reason is visible
Any 1 of these:
The order system or the payments system is unreachable or in a degraded state.
The carrier feed failed today.
The customer’s message mentions a chargeback, a complaint, a regulator or legal action.
The assistant cannot reliably match the customer request to the correct order.
A stop produces no payment write. It records the reason and routes the case to a named human: for example, “Credit not assessed: order system unavailable.”
Escalation path
Review and stopped cases go to the team lead’s queue, same day. If the queue already has 20 cases, the next case goes to the exceptions handler. If both are unavailable, the case waits without a credit being issued.
A timeout never turns it into an allowed action. A stopped case cannot resume until its stop condition is resolved.
Reversal and recovery path
Before enabling the allowed tier, the finance partner and team lead walk through a wrong-credit case using the actual payment system.
They establish which credits can be reversed, within what window, and who has permission to do it. Where reversal is impossible, the finance partner agrees the correction and the team lead owns the customer communication. Any generated customer message is reviewed before sending.
For this example, the team lead checks issued credits against the action log each working day, with alerts for failed writes and duplicate attempts.
Count wrong credits requiring correction, including those that cannot be reversed. If that count exceeds 5 in a month, suspend autonomous credits until the owners understand the cause and agree to restart. Stop conditions still apply while other cases move to review.
Log
Record every decision, including stopped cases: the order, proposed amount, evidence and timestamps, check results, boundary version, tier and reason, acting identity, human decision where applicable, payment outcome, and correction status. Give each attempted credit a stable identifier so a retry cannot quietly become a second credit.
The finance partner can read the log without an engineer. Agree access and retention with the owners under the organization’s existing records policy.
Signatures
Decision owner: the head of customer operations, for the boundary as written.
Resource owner: the finance partner, for the cap, the monthly correction trigger and the exposure.
Operating owner: the team lead, for the escalation path and the review queue.
Data owner: the platform team lead, for the carrier record and the 10-minute freshness rule.
The finance partner gets a limit and a recovery process to assess. The customer service rep gets a rule she can apply at 10:40 am. The architect gets the design inputs: which records to read, which identity writes, where it logs, and how it recovers.
The same 5 fields you already write
In September, I wrote about guardrails without gates: 5 fields, and the Tuesday Afternoon Test. Can a team apply this under deadline without booking time with you? This Action Boundary is that pattern applied to an AI action.
So if your org already writes guardrails, you already know how to write this. The new part is that the thing applying the guardrail on a Tuesday afternoon is software, potentially 2,000 times before lunch, which is why the stop conditions and the reversal path earn more attention than they would for a team of humans.
A guardrail is a decision that doesn’t need you in the room.
A gate is a decision you keep making.
Requiring approval for every AI action is a gate. As volume grows, the queue can erase the time saved, and approval can become a click without a check.
Your blank fill-in boundary
Take 30 minutes with the 1 action your org is debating.
Action. 1 sentence. What it does, to what, for whom.
Allowed. 3 to 5 conditions that must all hold. Use numbers: days, minutes, amounts, counts.
Review. 3 to 5 conditions, any 1 of which sends it to a human. Name the human’s queue.
Stop. 3 to 5 conditions, any 1 of which means do nothing and say so.
Escalation. Where review and stopped cases go, to whom, how long they may wait, and what happens if that person is away.
Recovery. How a wrong action is found, reversed or put right, who does it, how long it takes, and what correction trigger suspends the allowed tier.
Log. What’s recorded, who can read it without an engineer, how long it’s kept.
Signatures. Decision, resource, operating, data. Name the owner of each role.
Write the precedence above the cases: stop first, then review, then allowed. Decide what happens when evidence is missing or a service times out.
Then show it to the person in line 8 who’d carry the consequence. Tighten until they’d sign. If they haven’t heard of the proposal yet, you’ve found a real blocker.
________
🧭
________
Community Corner
Sunil Mishra brought the discussion back to recovery. He argued that authority, evidence, and recovery belong in the architecture from the start. In his follow-up, he emphasized what happens when an assistant gets something wrong: how it recovers is part of what makes it ready for production.
That gives us a useful test for the boundary above. Walk through the first wrong credit with the same care as the first correct one. Who notices, who acts, and what does the customer hear?
Help me make the next edition more useful.
I want to hear from more of you, including those who read each Sunday but haven’t replied before. Which part of this edition would you actually use, and what’s missing for the work you’re doing?
Hit reply to this email. One sentence is enough. If you’d rather discuss it with other readers, leave a comment when I share this edition on LinkedIn. A question, a disagreement or an example from your own work would all help decide what we need to explore next.
________
🧭
________
Where to start Monday
Pick 1 action. Write allowed, review, stop. Numbers, not adjectives. Show it to the person who’d get the call.
Next Sunday, closing this series: what would justify the next commitment? The sponsor is about to take “it saves 20 minutes a case” into a budget meeting.
Letting an AI system act is a different commitment from letting it draft. Bring the boundary above to the meeting with the people who would carry that commitment. Forward this to the person who would have to sign it.
A guardrail is a decision that doesn’t need you in the room.
A gate is a decision you keep making.
Until next Sunday,
Kevin
Work with me
I run Compass Architecture Advisory. I help enterprise and technology leaders turn AI ambition and architecture into decisions they can fund, govern and run.
Right now that’s a 4-week AI Architecture and Adoption Diagnostic. If there’s a decision like that in front of you, reply to this email with 2 lines about it, and I’ll let you know whether it fits.
________
________
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.
Kevin Donovan · Turning architecture into decisions
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.
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...
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...
Nov 1, 2026 | Read time: 10 mins. Kevin Donovan · Architects' HubForwarded this email? Subscribe to Architects’ Hub for the next issue. Where the weeks went Hey Reader, A warning arrives early. Permission arrives late. The recovery gets called heroic. Mike Carroll opens his article on decision latency with a composite plant-floor scene: a machine shows signs of trouble, but the person who can approve a controlled shutdown isn’t available. The line keeps running. Hours later, the bearing...