The State Between the Files

This is a follow-on to The Vendor Is Gone Friday. What Survives Monday?, which examined whether an organization can continue after changing AI providers. This piece asks what can still disappear even when every file survives the handoff. The scenario below is illustrative rather than a reported case. The broader argument and evidence are available in our Owned AI research briefing.

Lisa is the building maintenance coordinator at Auburn Heights. One morning she opens a work order and cannot tell what she is looking at.

A tenant on the fourth floor has reported water on their bathroom ceiling. Until a few weeks ago, a report like that would have reached Lisa already sorted, attached to an open work order with the last several reports from the same set of units. This one arrives as an emergency dispatch, ranked alongside a burst supply line.

Auburn Heights has an old riser that sweats through the cold months. Condensation collects where the insulation was patched badly during a renovation, and every winter it leaves a brown ring on the ceiling of the same three units. Tenants see a stain spreading above their heads and report a leak, which is the right thing to do. An engineer had already inspected the riser, opened a monitoring work order, and scheduled the insulation repair for summer.

The automated intake assistant read those reports the way it read any other report of water and sent them out as emergencies. After-hours contractors were driving across the city to photograph a stain that was already in the file, and the crew that should have been available for a genuine failure was standing in a fourth-floor bathroom looking at last month's ring.

So the team wrote an exception into the instructions the assistant loaded on every run. A ceiling stain in those units, during cold months, with no moving water and nothing wet below, gets attached to the monitoring order and scheduled during business hours. Anything else stays an emergency. Water someone can hear or see running, a stain in a new location, a light fixture involved, a ceiling beginning to sag.

It worked. Emergency dispatches at Auburn Heights dropped sharply, the contractor stopped billing for drives out to a known stain, and when genuine failures did occur, a crew was available that night. The written emergency policy never changed, because the team had not decided to ignore it. They had worked out how it applied to one recurring problem in one building.

Then the property manager changed software vendors. The migration carried the emergency policy, the inspection reports, the monitoring work order, and the complete history of every request the building had ever filed. The instruction that connected them did not come across in a form the new system could read, and nobody flagged it, because for two winters it had not been a rule anyone talked about. It was simply how the intake assistant behaved.

Now Lisa is looking at an emergency dispatch for a stain she is fairly sure she has seen before, and she has no way to confirm it. The obvious fix would be to tell the new system to stop treating ceiling stains at Auburn Heights as emergencies, which is roughly the right instruction with the safeguards filed off. The next tenant who reports a ceiling going soft above a light fixture would wait until Thursday.

What disappeared was not the emergency policy or the record of past repairs. It was the narrow judgment connecting them: which reports the policy covered, which it did not, and the conditions under which the exception stopped applying. That judgment had governed how an entire building got maintained until the system carrying it went away.

Four things a file can't tell you

Authority comes first. Lisa's team wrote the exception and it held for two winters, which is the only evidence anyone has that someone with standing approved it. A coordinator can reasonably decide how her own building's reports get sorted. Moving a habitability complaint from tonight to Tuesday may belong to the regional manager, and in some jurisdictions it belongs to whoever signs the compliance filing. The old system stored the instruction and the name of whoever typed it. It had no field for the question the new system needs answered, which is whether that person could give it.

Then there is commitment. When the assistant attached a report to the monitoring order, it told the tenant the stain was known, the repair was scheduled for summer, and someone would come by that week. That is a promise from the property manager, with a name on it and a clock running. The transcript migrated intact. It arrived as a sentence the old system once said rather than an obligation still owed to a fourth-floor apartment, so nothing in the new queue comes due.

Corrections create a different problem. The exception overrode the emergency policy, but only at Auburn Heights, only for the three units on that riser, only in cold months, and only while no water was moving. Each of those limits got added after some report tested the previous version of the rule. The workspace kept the final wording and lost the shape: which policy it bent, how far, and the month it was supposed to stop applying. A successor reading that instruction in July would enforce a rule built for February.

The fourth is outcome. Somewhere in two winters of downgraded reports is the answer to whether the exception was sound. Most of those ceilings were the same seasonal ring, dry again by March. One of them may have been a slow supply line inside a wall, found months later and written up as its own emergency, with nothing linking it back to the morning the rule sent it to the business-hours queue. The reports survived the migration and so did the repair history, in separate systems. What did not survive is the line between them that would tell a successor whether it inherited a judgment that worked or a gamble nobody ever settled.

None of these four is context in the ordinary sense. Context is background that helps a system give a better answer. These four decide whether an action is allowed at all. A successor that guesses wrong about them doesn't produce a slightly worse answer. It produces an action the institution itself would not recognize as valid, and it does so confidently, because everything it can see tells it the written policy is the rule.

The portable pieces are arriving

Procedure can travel. Anthropic's Agent Skills package a workflow, its scripts, and its reference material into a versioned folder that works across several tools. The Model Context Protocol gives models a standard way to reach tools, data, and prompts, reducing the proprietary plumbing between an assistant and the systems it drives.

Google's Open Knowledge Format does something similar for curated organizational knowledge. The deliberately minimal specification uses Markdown files with YAML headers, with no central registry or required tooling. It standardizes enough to make a body of knowledge self-describing: what each concept is, where it came from, and how it links to the rest.

Memory can also sit in customer-controlled storage. Anthropic's memory tool reads and writes against infrastructure the customer operates. Microsoft's Work IQ takes the managed route, combining organizational data, persistent memory, and inference under scoped permissions and policy enforcement, with traces of what an agent called and what came back.

Those formats carry useful parts of the building's record, but they do not describe which rule currently governs. At Auburn Heights, the exception superseded the emergency policy for three units during cold months, while the engineer's inspection report merely related to the same building and riser. The Open Knowledge Format can capture both facts, cite their sources, and link each to another concept. Because those links carry no type, "supersedes" and "relates to" look identical to the system reading them. The file can move the knowledge without carrying the decision that this correction overrides that rule, for these cases, on this person's authority.

Work IQ carries permissions and policy enforcement inside Microsoft's control plane. Those controls can decide what an agent may open or invoke. The institution's decision rights answer a different question. They determine whether the person behind the agent may waive a policy, promise an outcome, or delegate that power. A successor can inherit access to a customer record without inheriting the authority that governed how the record was used, and whatever authority the original control plane did encode may be difficult to reconstruct outside it. The pieces are converging quickly around a seam that no format or control plane has claimed.

OpenAI's own documentation illustrates the gap between holding data and resuming work from it. ChatGPT Business offers no settings-based export. Merging a personal workspace into one deletes its custom instructions and plugins, while a new Business workspace inherits nothing from a deactivated one. The data can remain intact inside the old workspace without leaving in a form that lets another system resume the work.

We have watched this fail before

NASA ran an agency-wide Lessons Learned Information System that managers were required to review, supported by center-level and program-level systems. In 2002, the Government Accountability Office found no assurance that the lessons were being applied. Program managers were not routinely identifying or sharing them, and the barriers were cultural as much as technical. People lacked time, and many believed mistakes were not safe to admit. Formal systems and a written mandate had not made institutional memory part of the work.

A better schema would not have fixed NASA's system. The failures ran through capture, retrieval, sharing, and application. An agent can change one link in that chain without repairing the rest.

The old failure ran partly through the reader. Consulting the record required a busy person, under deadline, to remember that a relevant lesson might exist, go looking for it, and choose to act on it before doing the thing they were already going to do. Every step in that chain leaks, and the leaks compound. An agent that loads its operating context before it acts can be configured to consult the approved record every time, without depending on anyone to remember the record is there.

Reliable retrieval leaves the other failures in place. Capture still competes with the day's deadlines, records still go stale or lose the conditions that made them true, and the cultural problem the GAO found remains untouched. People do not document mistakes in an organization that punishes them for doing so, and a flawless reader cannot repair a record no one was willing to write. Agents make approved records easier to deliver at the moment of action. That is a narrower improvement than institutional memory, but it moves one part of the bottleneck to a place with better tools than NASA had in 2002.

The version of this that becomes surveillance

The same record that helps a successor resume the work can also help a manager, insurer, or opposing counsel inspect the people who did it. Capturing decisions, corrections, authority, and outcomes at that level can turn continuity infrastructure into a permanent behavioral file.

An OECD survey of more than 6,000 firms across France, Germany, Italy, Japan, Spain, and the United States found algorithmic management already widespread. Managers reported unclear accountability and tools whose logic they could not follow. In the United States, 55 percent of firms monitored the content and tone of conversations, calls, or emails, compared with roughly 6 percent in the European countries surveyed and 8 percent in Japan. Across firms where managers reported data collection, 65 percent said workers could not opt out, 22 percent said workers could not access their data, and 30 percent said they could not request corrections.

A continuity record should be narrower than a memory archive. It needs the approved interpretation, the scope of an approval, an open commitment, the correction that displaced an older rule, and evidence of what actually happened. Drafts that never governed a decision, private commentary, speculative judgments about workers, and every chat message kept as insurance belong outside it by default.

A governing record with no accountable owner, expiration, or usable correction path becomes a file on people who cannot see or challenge it. Those controls are part of the continuity mechanism because they determine when a record still governs and who can correct it.

An institution should survive changing its mind

Hand a replacement system your continuity package. Reconnect the approved sources and tools, and nothing else, then ask it to resume one live piece of work. A successor that has what it needs can take the next action in a way the people who set the rules would recognize as valid.

For Auburn Heights, that package is small. It records who carved the exception and how far their authority reached, what tenants were told to expect, which written rule the correction overrode, and how the downgraded reports turned out. A different model working from those records can sound nothing like the old assistant and still route the next ceiling stain correctly, because continuity shows up in the resumed work rather than in an imitation of the old system's prose.

The industry is pouring enormous effort into teaching AI products to remember our preferences, habits, and histories. Institutions face the inverse problem: remembering themselves without depending on any one product to do it for them. That means preserving enough of their decisions, commitments, corrections, and outcomes for the next system to pick up the work without quietly changing what those decisions meant.