When the Agent Goes Down, Who Can Still Do the Work?

Published: at 12:50 AMSuggest Changes

The agent stopped accepting work at 02:40 and the escalation landed with a duty manager who had never processed one of these cases by hand. The runbook said revert to manual. Manual was a retired queue, a spreadsheet nobody maintained, and a procedure last rehearsed the week before the agent went live. The agent came back at 06:15, and the team spent the intervening hours waiting, escalating, and improvising a workaround that will not survive contact with an audit.

Two failure modes get planned for when an organisation deploys an agent. The agent stops, or the agent is wrong. Both eventually put a human in front of a decision. A third rarely gets a meeting of its own: the human path was removed while everyone watched the agent’s uptime.

Availability is getting worse, not better

The platforms agents run on have become less reliable as adoption has grown. Ookla analysed 471 days of US Downdetector data from January 2025 to mid-April 2026, covering 3.72 million user-reported problems, and counted days when ChatGPT, Claude, Gemini or Microsoft Copilot drew more than ten times their normal report volume. The count went from six in the first quarter of 2025 to sixteen in the fourth, then fifty-one in the first quarter of 2026. [Ookla, June 2026]

This is no longer an early-adopter problem. Open Future Forum’s October 2026 AI Leaders survey found 80% of respondents running at least one production agent and 29% running more than twenty, on a selective, self-reported sample of 75. Even discounted, the direction is clear: production agents have become ordinary, and ordinary things get relied upon without anyone deciding to rely on them. [Open Future Forum, October 2026]

The most instructive event was not an outage. On 12 June 2026 the US government applied export controls to two Anthropic models, Fable 5 and Mythos 5. Because the order took effect immediately and nationality could not be verified in real time, Anthropic suspended access for every user worldwide, restoring it on 1 July. [Anthropic, July 2026]

Nineteen days, one vendor, no notice, and no defect to fix. A plan that treats a second model from the same provider as failover does not survive that.

Singapore’s agentic AI guidance has now named the consequence. The Model AI Governance Framework for Agentic AI, updated to version 1.5 in May 2026, tells organisations that as agents take over entry-level tasks — the work that traditionally trained new staff — users can lose basic operational knowledge, creating a business continuity risk in which they may no longer know how to perform critical processes manually when agents malfunction or become unavailable. The instruction is specific: identify the core capabilities of each job and give people enough training and work exposure to retain the foundational skills. [IMDA, May 2026]

That is a governance document telling you your continuity plan has a staffing clause.

The skills that fade are the ones an incident needs

The evidence on what happens to human capability when automated systems absorb routine work is more specific than “people get rusty”.

Research on customer support agents found that a generative AI assistant raised productivity by around 15%, with gains concentrated among less-experienced workers, and helped newer workers move down the experience curve. [Brynjolfsson, Li and Raymond, QJE, May 2025] The assistant was strongest on the tasks a junior learns on. Take those out of the job and the apprenticeship that produced your current experts stops producing them.

Hiring is moving the same way. Stanford’s Digital Economy Lab has tracked employment for 22 to 25 year-olds in AI-exposed occupations, and its August 2026 update puts that group 19% below where it would be had it kept pace with less-exposed peers. The authors call this descriptive evidence rather than a causal estimate, and I would read it the same way. Hire fewer juniors into an AI-exposed function while automating the tasks they would have learned on, and you are deciding where your next generation of incident-capable operators comes from. [Brynjolfsson, Chandar and Chen, August 2026]

Aviation holds the longest record here: the industry automated the same way, then studied the result. A NASA-led study put airline pilots through a 747-400 simulator, varied the level of cockpit automation, then tested what happened when it was removed. Instrument scanning and manual control skills held up reasonably well, even among pilots who practised them rarely. The cognitive skills did not: tracking position without a map display, working out which navigational step came next, and recognising instrument system failures all produced frequent and significant problems, and the decline tracked how often pilots let their attention drift while the automation ran. [Casner, Geven, Recker and Schooler, Human Factors, 2014]

Your operators will keep the ability to press the buttons; procedural memory is durable. What atrophies is the ability to hold a process in your head, work out what comes next, and diagnose a failure you have not seen before — under time pressure, at 02:40, with customers waiting.

A fallback is a capability, not a paragraph

Most organisations have a written fallback. Far fewer have a capability, and financial services has the vocabulary for the difference.

The FCA’s March 2026 review of operational resilience restates the expectation: firms identify their important business services, set impact tolerances, map the dependencies, and test whether they can remain within tolerance through severe but plausible disruptions. [FCA, March 2026]

An agent becoming unavailable for a fortnight is a severe but plausible disruption. The test is whether the recovery option has been exercised, by the people who would run it, at the volume required. A recovery option resting on unnamed staff who have not performed the task since last year’s exercise is not a tested option, whatever the document says.

Three things keep a manual path alive, and all three are operational.

A proficiency standard with named holders. Define the minimum: this task, to this quality, at this rate. Then name the people who meet it today, because a rota is what you will reach for during an incident.

A declared trigger and a decision owner. Degraded mode needs an invoker, with a threshold that is not “when it feels bad”, and someone must own the call to return to normal. Ambiguity costs hours, and hours are what an impact tolerance is denominated in.

A rehearsal cadence that lives in operations. An annual walkthrough in a meeting room does not maintain a skill. Working the manual path at low volume, on a schedule, does.

One further signal is worth watching while the agent is up. The IMDA framework suggests tracking the human override rate — how often reviewers reject or modify what an agent proposes — and response times during review, on the basis that a very low override rate may signal rubber-stamping. If your override rate has drifted towards zero and review times keep falling, the signature is still on the decision and the capability behind it is not. [IMDA, May 2026]

Design the degraded service

Sort the work. Some processes can absorb delay — queue it, batch it, tell the customer it will take longer. Some can run in a thinner form: narrower scope, lower service level, a manual check where an automated one used to sit. A few have to keep moving, and those need a staffed, trained and measured human lane.

Then keep that lane warm. Route a deliberate share of live volume through the manual path permanently, and accept the throughput cost. It is the only method I have seen that keeps a manual capability honest: the people running it work on real cases with real deadlines rather than a script in a drill. Airlines reached the same answer — manual flying on some sectors, recurrent simulator checks — and even that discipline only partly protects the cognitive skills. A paragraph in a continuity plan protects nothing.

Two rules follow. Keep the fallback off the primary’s dependencies: a second model from the same provider is not failover against a provider-wide or regulatory failure, and a manual path running inside the same ticketing system or identity provider the agent depends on is the same outage with more steps. Then state numbers rather than intentions. “We revert to manual” tells a board nothing; “we can process 40 an hour by hand against a normal 400, which holds the service inside tolerance for about six hours” is a plan with an expiry, and it forces the staffing conversation the intention was avoiding.

The evidence to require before the human path is removed

If you are about to take the human out of a critical process, or reduce the team behind it, the following should exist and be current. Most of it is uncomfortable to produce.

That last item is where programmes blink. If the manual path cannot be made viable, redesign the tolerance, the process or the staffing rather than keeping a control on paper nobody can operate.

The test

Pick the three processes where an agent outage would breach your tolerance. Put them to the accountable owner and ask three questions. When was this last run end to end by hand? How many can we process per hour that way? Who is on today’s rota who has done it?

If the answer to the first is last year, the second arrives as a range, and the third requires a search, you do not have a fallback. You have an assumption, and the agent’s uptime is the only thing holding it up.

Sources

Previous Post
Agents Are Software: Give Them a Delivery Pipeline
Next Post
Explainability Under Pressure: AI Decisions in Credit, Insurance and Hiring