Six months after most AI consulting engagements end, one of three things has happened to the AI system that was built.
The first: it is running well and improving. The team is using it consistently, the AI system owner is maintaining it, and the outputs are better than they were at engagement end.
The second: it is running but not improving. The team is using it, but the context pack has not been updated in four months and the acceptance rates have declined slightly.
The third: it has quietly stopped being used. Two team members still open it occasionally. Everyone else reverted to their previous approach within eight weeks of the engagement ending.
The outcome is determined almost entirely by decisions made during the engagement, not after it.
The system that compounds after the consulting firm exits was built with the structural elements that produce compounding. The system that degrades was built without them.
This article describes the structural difference and what a founder can do, during the engagement, not after, to determine which trajectory the system is on.
The three post-engagement trajectories: what each looks like
Trajectory 1: Compounding
At month 6 post-engagement:
The AI system owner is running the weekly maintenance cadence consistently. The blended acceptance rate is 3 to 5 percentage points higher than it was at engagement end because 24 improvement cycles have been run since the consulting firm exited.
Two new workflows have been documented and deployed by the AI system owner independently. One team member who was not using AI at engagement end is now using it consistently, after the system owner ran a targeted anchor workflow session.
At month 12 post-engagement:
The AI system is materially better than it was at engagement end. The context pack has been through eight quarterly review cycles.
The workflow library has grown from five to eight workflows. The blended acceptance rate is 88%, eight points above the 80% target reached at engagement end.
The AI system owner handles maintenance independently. The consulting firm has been consulted twice for strategic guidance: once for adding a new workflow cluster for a new service line, once for evaluating a new AI tool for Phase 3.
What produced this trajectory:
- A capable AI system owner with real time allocation (5 hours per week) trained to independence before the consulting firm exited
- A context pack built with deep enough founder input to remain accurate for 6 to 12 months before needing significant update
- Workflows built with documented specifications that the system owner can improve without rebuilding
- A monthly deliverable agreement for the first 90 days post-engagement that established the maintenance habit
Trajectory 2: Plateau
At month 6 post-engagement:
The AI system owner is doing the minimum viable maintenance. The adoption log is being reviewed but the reviews are not producing systematic improvements.
The context pack has been updated once (a pricing change in month 2), but the three new service offerings launched in months 3 and 4 have not been added.
The blended acceptance rate is roughly where it was at engagement end, 77%, neither improving nor declining significantly. The team is using AI consistently for the three highest-adoption workflows. The two lower-adoption workflows have not grown.
At month 12 post-engagement:
Context decay has begun to show in output quality on the newer service offerings. The AI outputs reference old service descriptions. Two team members who were on the adoption boundary at engagement end are now using AI inconsistently.
The acceptance rate on the two lowest-performing workflows has declined from 72% to 65% because no improvement cycles have been run on them.
The system owner is aware the context pack is out of date but has not had the time to update it.
What produced this trajectory:
- An AI system owner who is technically capable but whose time allocation was cut when the engagement ended
- A context pack that was accurate at engagement end but required monthly updates to stay current
- No formal post-engagement accountability structure (quarterly reviews, monthly deliverable agreements) to prevent gradual drift
The plateau trajectory is not a failure. It is the baseline. The risk is that “maintained without improvement” often means “gradually declining,” and what looks like maintenance is actually slow decay.
Trajectory 3: Decay
At month 6 post-engagement:
The AI system owner role effectively ended when the consulting firm exited, not through a decision but through displacement.
The person named as AI system owner returned to their full operational role and the 5 hours per week that were protected during the engagement were immediately consumed.
The adoption log has not been reviewed in eight weeks. The context pack has not been updated since the consulting firm’s last session.
Two of the five deployed workflows are producing outputs the team has stopped trusting. One was a client communication workflow that produced an output with an outdated pricing reference. The team stopped using it and reverted to manual.
At month 12 post-engagement:
Two team members are still using Claude for individual tasks in their personal accounts. The shared workspace is technically accessible but unused. The context pack loaded into the shared workspace is nine months out of date.
The AI system the company paid $45,000 to build is producing the same outputs as a generic Claude account, because the context that made it company-specific has not been maintained.
What produced this trajectory:
- An AI system owner who was named but not given protected time
- No monthly deliverable agreement or post-engagement accountability structure
- The consulting firm’s exit timed to the calendar rather than the system state, and the system owner was not independent before the exit
The five structural conditions that determine trajectory
Condition 1: A trained, independent AI system owner with protected time
The verification test: two weeks before the engagement end, the AI system owner runs the weekly maintenance cycle independently, reviewing the adoption log, identifying the improvement action, executing the update, and validating the improvement, without any consulting firm involvement. The consulting firm observes but does not assist.
If the system owner completes this cycle correctly, independence is demonstrated.
The most common failure mode: the system owner completes the cycle with consulting firm assistance during the engagement but has not been given a solo cycle before exit. The first solo cycle after exit is also the first time the system owner is on their own, which is when gaps in capability become apparent.
Condition 2: A context pack that is current at engagement end
The verification test: on the last day of the engagement, the context pack is reviewed against the current business state.
- Are all client archetypes accurate?
- Do the service descriptions reflect current offerings?
- Are the decision rules current?
- Is the competitive positioning accurate?
Any gap identified at this review should be closed before the engagement ends, not noted for the system owner to address later.
The most common failure mode: the context pack was accurate in month one but has not been systematically updated as the business changed during the engagement. The system owner inherits a context pack that is already partially outdated at handover.
Condition 3: A monthly deliverable agreement for the first 90 days post-engagement
The verification test: before the engagement ends, the consulting firm and the company agree on specific outputs for months one, two, and three post-engagement. The agreement is written and signed. The consulting firm is available for a monthly 30-minute review call to assess whether the agreed outputs were produced.
Why the first 90 days are critical: the maintenance habit forms or breaks in the first 90 days after the consulting firm exits. A system owner who completes the first three monthly cycles with accountability develops the habit. One who has no accountability structure in the first 90 days typically does not.
The form of the agreement:
POST-ENGAGEMENT 90-DAY DELIVERABLE AGREEMENT
----------------------------------------------
Month 1:
- Context pack reviewed, with all gaps from the final engagement review addressed
- Two adoption log review cycles completed independently
- Report: blended acceptance rate and adoption frequency by team member
Month 2:
- One workflow below threshold diagnosed and improved
- Context pack updated with any business changes from month 1
- Report: acceptance rate change on improved workflow
Month 3:
- Quarterly health check conducted
- Results presented to founder at quarterly review meeting
- One new workflow candidate identified for potential Phase 3Condition 4: A quarterly founder review of AI system health
The verification test: the founder’s calendar has a 45-minute quarterly AI system review blocked for the next four quarters before the engagement ends. The agenda is specific:
- Blended acceptance rate review
- Adoption frequency by team member
- Context pack currency check
- One strategic question: what is the highest-value workflow we should add next?
Why the quarterly review matters: it provides the accountability structure that converts the AI system owner’s maintenance work from “something I try to do” to “something I report on.” The system owner who knows the founder reviews the health check quarterly is more likely to maintain it than one who reports to no one.
Condition 5: A documented escalation path for situations the system owner cannot handle
The verification test: the AI system owner can name three things: who to contact at the consulting firm for escalations they cannot handle, what the response time expectation is, and what the consulting firm charges for post-engagement support.
The most common gap: the consulting firm exits without agreeing on post-engagement support terms. The system owner encounters an issue they cannot handle (a new tool integration, a significant context pack restructuring, a team expansion requiring new role-specific workflows) and has no clear path to get help without re-engaging the full project process.
The form: a post-engagement support agreement. Typically a 3-hour bank of consulting hours per quarter at the engagement’s hourly rate, available for escalations only. Not a retainer. A safety net.
What to do if the engagement has already ended and the system is decaying
Early-stage decay (2 to 4 months post-engagement)
Signs: context pack slightly outdated, adoption beginning to decline.
Recovery: a one-week internal sprint. The AI system owner runs a context pack audit, reviewing every entry against current business reality and updating what has changed. They then run three adoption sessions with the team members whose usage has declined, using the anchor workflow approach on real current tasks.
Improvement is visible within two weeks. No consulting firm involvement required.
Mid-stage decay (5 to 8 months post-engagement)
Signs: context pack significantly outdated, 30 to 50% of team has reverted.
Recovery: a targeted re-engagement, not a full Phase 1 rebuild. A focused 2 to 3 week context pack update and adoption reset.
The consulting firm or the AI system owner (with consulting firm oversight) conducts a context pack audit, rewrites the sections that have drifted, and runs a second round of anchor workflow training for the team members who have reverted.
Total cost: $5,000 to $10,000. Timeline: 3 weeks.
Late-stage decay (9 or more months post-engagement)
Signs: system essentially unused, context pack completely outdated.
Recovery: a full Phase 1 rebuild, but starting from a position significantly better than zero.
- The original workflow specifications still exist, even if they need updating
- The team has AI familiarity from the first engagement
- The AI system owner knows the maintenance cadence even if they have not been running it
- Phase 1 rebuild takes 2 to 3 weeks instead of 4 to 6 weeks
- Phase 2 training takes half the time because the team is not starting from scratch
Cost: $15,000 to $25,000 for a targeted Phase 1 rebuild and Phase 2 reset.
The lesson from late-stage decay: the investment in the original engagement was not entirely lost. The team’s AI familiarity, the original workflow specifications, and the institutional knowledge of what worked and what did not are all preserved. They make the second engagement faster and cheaper than the first.
The most expensive outcome is not the failed engagement. It is waiting 12 or more months to address the decay, during which time the competitive gap widens.
Common questions on post-engagement sustainability
”How do I verify the five structural conditions before signing an engagement?”
Ask five specific questions before signing.
- “Will the AI system owner have run a solo maintenance cycle before you exit?”
- “Will you conduct a context pack currency review on the last day of the engagement?”
- “Will you provide a written 90-day post-engagement deliverable agreement?”
- “Do you recommend a specific quarterly founder review format?”
- “What are your post-engagement support terms for escalations?”
A firm that answers all five with specifics is designed to produce the compounding trajectory.
A firm that answers vaguely or asks “do we need to include that?” is designed to produce a deliverable at engagement end, not a system that compounds after.
”Is it normal for the AI system to decline slightly after the consulting firm exits before it stabilises?”
A small dip in the first 4 to 6 weeks post-engagement is normal as the AI system owner takes over fully and the team adjusts to the consulting firm’s absence.
Acceptance rates may drop 2 to 4 points temporarily before stabilising.
A decline of 10 or more points over the first 90 days, or a decline that continues without stabilising after 8 weeks, is not normal.
It indicates either a system owner capacity problem or a context pack that was not current at handover.
”What is the minimum time the AI system owner needs to be in place before the engagement exits?”
The AI system owner should be named at the engagement’s start and actively shadowing the consulting firm’s maintenance work by the end of Phase 1 (typically weeks 4 to 6).
They should be taking the lead on maintenance by the midpoint of Phase 2 (typically month 3). They should be fully independent, with the consulting firm observing only, in the last two weeks before engagement exit.
An engagement that names the system owner in the final two weeks has not developed the system owner. It has informed them of a role they are inheriting without preparation.
Want an engagement structured so the system compounds after we leave, not one that peaks when we exit?
What happens after the AI consulting engagement ends is the most important question most founders never ask before they sign.
The answer depends almost entirely on decisions made during the engagement: whether the AI system owner was trained to independence, whether the context pack is current at handover, whether a post-engagement accountability structure was established.
And whether the consulting firm’s exit was timed to system state rather than calendar.
The three trajectories, compounding, plateau, decay, are set before the engagement ends. The structural conditions that determine which trajectory the system is on are specific and verifiable. Checking them before the consulting firm exits is the most important investment the founder makes in the engagement’s lasting value.
Path one: verify the five structural conditions at your next engagement review meeting. Use the five verification tests above as your agenda. Any condition that cannot be verified is a specific gap to close before the engagement ends.
Path two: bring in a partner. The five structural conditions in this article are the design principles of every Phos AI Labs engagement: the system owner trained to independence before exit, the context pack reviewed and current at handover, the monthly deliverable agreement for the first 90 days, the quarterly founder review calendar blocked, and the post-engagement support agreement in place. We have run 400+ AI engagements. Clients include Zapier, Coca-Cola, Medtronic, Dataiku, and American Express. Thirty minutes, no deck. Start here.