Every family has a basement it plans to clean. The plan is always total. One weekend, everything comes out, gets sorted, gets labelled, goes back in order. The weekend never arrives, because the job is too big to start and too vague to finish. Meanwhile the family keeps living upstairs, and when someone needs the camping stove they go down, move four boxes, find it, and come back up. Nobody ever needed a clean basement. They needed the stove.
Most companies treat data readiness for AI adoption like that basement. October 2026 brought a pile of surveys showing how few of them have done the cleanup, and the numbers are bad enough to make any board nervous. My argument is that the numbers are measuring the wrong unit. Enterprise-wide data readiness is a weekend that never comes. The companies getting value from AI this year went downstairs for one thing, found it, and wired it to one workflow. The surveys can't see what they are doing, because the surveys ask about the whole basement.
What October's numbers actually say
Start with the receipts, because they are striking.
On October 13, the data and AI services firm Further released its State of Enterprise AI Maturity report, based on 515 U.S. AI decision-makers. Pure AI reported on October 19 that just 5% believed their organization's data was largely ready for AI. In the previous 12 months, 37% had paused, scaled back or abandoned an underperforming AI initiative, and 24% named data quality or availability as the main reason for the setback. Only 29% had completed a formal data audit and remediation plan.
Then the twist. In the same sample, 77% reported a positive return on AI investment.
Hold those two figures side by side. Almost nobody thinks their data is ready, and three quarters say they are already making money. If data readiness were the precondition everyone says it is, those numbers could not both be true. They are true, so it isn't the precondition, at least not in the form the survey asks about.
Komprise found the same shape in unstructured data. Blocks & Files covered its survey on October 19. Sixty-eight percent of respondents call their organization mature or maturing on AI. Seven percent have classified and cleaned most of their file data. Four percent have developed schema for most file data. Fifty-seven percent say half or less of their data is ready, and 13% cannot say how much is ready at all. Komprise COO Krishna Subramanian put it gently: "Enterprise IT and AI leaders recognize that data intelligence is lacking but the pathway to ensuring high data quality is not straightforward."
OutSystems added a third angle with research run alongside CIO Dive and KPMG, reported by iTWire on October 19. Of more than 850 IT decision-makers, 82% believe their organizations can keep pace with governing, integrating and monitoring AI agents. Twenty-six percent say they are well prepared to resolve conflicts across agents, systems and workflows. Thirty-five percent cite data consistency and context sharing as a major orchestration hurdle. Five percent have automated controls, audit trails and formal human review points in place.
Read in sequence, the three surveys tell one story. Confidence is high. Measured readiness is tiny. Returns are arriving anyway.
The vendors read the same numbers
The market has noticed. On October 6, Precisely announced the Precisely Platform, a unified data management product built on one metadata and semantic foundation, covering integration, quality, governance and master data, with quality and governance scores on every data asset. It ships Model Context Protocol servers so that Claude, Microsoft Copilot and ChatGPT can reach it directly. It is in preview, with general availability slated for early 2027.
On October 19, Boomi announced it had acquired Elementary Data, a Tel Aviv data observability company that started in 2022 as an open-source project and is used by more than 1,500 data teams. Boomi says the deal brings column-level lineage, pipeline monitoring for freshness and volume anomalies, health scores, and an MCP server that exposes data context to AI tools and agents. Boomi CEO Steve Lucas summed up the pitch: "AI is only as trustworthy as the data behind it." The announcement cites a Forrester Consulting study, commissioned by Boomi in July 2026, in which organizations that deployed agents before they were ready reported an average of $2.1 million in added cost.
That $2.1 million figure deserves the skepticism any vendor-commissioned number deserves. The product design deserves more attention than the number does. Both Precisely and Boomi built MCP servers. Both attached scores to individual assets. Neither one is selling a finished basement. They are selling a way for a machine to ask, at the moment it needs something, whether that particular box can be trusted.
The reader changed
For twenty years, the reader of enterprise data was a person with a dashboard. Data readiness meant the warehouse was tidy enough that an analyst could build a report without lying by accident. The unit of readiness was the warehouse, because the analyst might ask anything of it, and you could not predict which question would come next.
That reader is being replaced, and October made the replacement concrete. On October 8, Google introduced a unified agent in Gemini for businesses. TechCrunch reported that it connects to Google Workspace, Microsoft 365, Slack, Jira, Confluence, Git, BigQuery, Databricks, Postgres and Snowflake, works with any MCP server inside or outside a company's network, and gets its own Workspace account, with an email address and context about colleagues, approvers and calendars. It writes audit trails attributed to the agent rather than a person. Sundar Pichai said the business-first rollout lets Google address the "harder problems around security, scale, and performance" that come with powerful agents.
An agent reads differently from an analyst. An analyst wanders. An agent assigned to a task reads a specific, traceable set of fields, in a specific order, to do a specific thing. Its appetite is narrow and its footprint is logged. When it fails, it fails at a particular join, on a particular column, with a particular stale value, and the trace tells you which one.
The narrowness is the opening. When the reader was a wandering human, enterprise-wide readiness was the only sensible target, since you couldn't know where they would wander. When the reader is an agent with a job, readiness can be scoped to the job. The data the agent touches for invoice matching might be a few dozen fields across three systems. You can get those right in weeks. You will never get the whole estate right.
Why the surveys keep failing everyone
The survey question "is your data ready for AI?" assumes one answer for the whole company. It has to be answered by someone who is thinking about the whole basement, which means the honest answer is always no. Ask the same executive whether the data behind their returns-authorization workflow is ready, and you will get a much more useful answer, possibly yes.
So the 5% figure from Further and the 7% figure from Komprise are accurate and nearly useless as operating guidance. They describe a state almost no company will ever reach and almost no company needs to reach. The 77% positive-return figure, sitting in the same Further sample, is the companies that quietly stopped waiting.
Readiness is a contract, signed per workflow
Here is the position I will defend. Kill the enterprise data readiness program as a gate. Replace it with a readiness contract for each workflow you put an agent on.
A readiness contract is short. It names the decision the agent makes. It lists every field the agent reads to make that decision, with the system of record for each. It sets a freshness tolerance for each field, a known error rate, and the owner who gets paged when either drifts. It says what the agent does when a field fails the check. Usually that means stop and hand the case to a person.
Suppose the workflow is supplier invoice exceptions. The contract might cover the purchase order number, the receiving record, unit price, quantity, payment terms and the vendor master entry. Suppose that comes to around forty fields across an ERP, a warehouse system and a vendor portal. A team can profile forty fields, find the duplicates in the vendor master, fix the unit-of-measure mismatches, and wire freshness checks in a sprint or two. The other several thousand tables in the warehouse stay exactly as messy as they were, and nobody cares, because the agent never reads them.
This is what the Boomi and Precisely products are quietly built for, whether or not their marketing says so. A health score on one asset, exposed over MCP, is a contract clause a machine can check at runtime. Column-level lineage tells you which upstream job broke your forty fields. The tooling has moved to the workflow unit. Most governance programs are still organized around the estate.
What the contract changes inside the company
Three things change when you work this way.
First, ownership gets real. An enterprise data quality program has a sponsor and a steering committee, and no single person loses sleep when the vendor master has 11,000 duplicates. A workflow contract has a named owner for each field, and that owner knows exactly which agent breaks when their field goes stale. Accountability attaches to something small enough to hold.
Second, the budget argument flips. A request to clean the data lake for AI has no natural end and no natural return, which is why CFOs fund it at half strength and then cut it in the next downturn. A request to fix forty fields so that the invoice agent can clear exceptions without a human has a cost, a date and a payback. Further found that only 29% of respondents had completed a formal data audit and remediation plan. I suspect the other 71% are mostly not lazy. They were asked to fund a basement.
Third, the failures become your map. Every time an agent stops on a field check and hands a case to a person, you have learned something precise about where your data is weak, in the place where weakness costs money. After six months of contracts across eight or ten workflows, you will know more about the condition of your data than any enterprise audit could tell you, because the audit sampled everything and the agents tested what matters.
The objection worth taking seriously
The strongest pushback goes like this. Workflow-by-workflow readiness builds silos. Each team cleans its own forty fields, defines "customer" its own way, and five years later the company has twelve incompatible versions of the truth, each one tidy on its own, and a worse mess than before. Master data management exists precisely because local fixes drift apart.
The objection is correct about the risk. It is wrong about the remedy. The answer to drift is a shared definition layer that the contracts point to, built up as contracts demand it. The first time two workflows both read the vendor master, that table gets one owner and one definition, and both contracts reference it. Shared entities emerge from overlap, in the order the business needs them, instead of being specified up front by a committee guessing which entities will matter.
Precisely's design is a fair example of where the market is going. Its platform works across data where it already lives, without requiring migration, on a shared metadata layer. That is a semantic spine you can grow incrementally. It does not require the whole basement to be emptied first.
There is a second objection, quieter and more dangerous. Agents with broad access, like the Gemini agent with its long list of connectors, will read beyond the contract. Give an agent an objective and access to Snowflake, and it may decide on its own that some unvetted table is relevant. The OutSystems figure matters here. Only 5% of respondents have automated controls, audit trails and formal human review points in place. If the contract is a document and the agent's permissions are broad, the contract is decoration.
So the contract has to be enforced in the access layer. The agent for invoice exceptions gets credentials scoped to the fields in its contract and nothing else. If it needs a new field, that is a contract amendment with an owner, a freshness check and a review. Breadth is the default in most agent products shipping this autumn, because breadth demos well. Narrow scoping is your job. Kierra Dotson at Further put the general point well: "confidence without a documented process is exposure."
What to do on Monday
I'd rather give you a sequence than a framework.
Pick the one workflow where a human spends the most hours moving information between systems that already hold most of the answer. Exceptions, reconciliations, intake triage and claims first-notice are the usual candidates. Write down the decision in one sentence. Trace every field a competent person reads to make that decision, and write the list on a single page. If it runs past two pages, the decision is too big; split it.
Profile those fields this month. Measure null rates, duplicates, staleness and disagreement between systems. Fix what you can, and for what you can't, write the rule the agent follows when it meets the bad value. Then scope the agent's credentials to exactly that list, turn on the field checks, and let it run with a person reviewing its handoffs.
After the first workflow, do the second. When the second workflow reads a table the first one already touched, you have found your first shared entity. Give it a single owner. Keep a running count of how many fields across the company are under contract. That number, rising quarter by quarter, is a readiness metric that means something, unlike a survey answer about the whole estate.
Then stop answering readiness surveys with a shrug. When a board member asks whether the company's data is ready for AI, the right answer is a list: these six workflows, these 300 fields, these owners, this exception rate, trending down. The question as asked has no good answer. The list does.
What this costs and what it saves
Suppose a mid-sized manufacturer had budgeted an eighteen-month enterprise data quality program before deploying any agents. Under the contract approach, the first agent goes live in a quarter, funded from the same budget, and the remaining money gets spent fixing the fields that the first several agents proved were broken. The total spend may end up similar. The order changes everything. Value arrives in the first quarter instead of the sixth, and the cleanup is aimed by evidence instead of by a consultant's inventory.
The $2.1 million average from Boomi's commissioned Forrester study is the cost of skipping readiness. The cost of chasing total readiness never shows up in a survey, because the companies paying it haven't shipped anything yet. Further's 37% who paused, scaled back or abandoned an initiative contain both kinds of failure. Some went too fast without contracts. Some were waiting on a basement.
The conclusion is a decision
Every month you leave data readiness framed as an enterprise program, you are choosing a project with no end date over a set of projects with payback dates. The surveys published this October will be quoted in your next budget meeting as proof that the company isn't ready. Treat them as proof that the question was badly asked.
A tool will not fix this for you. Precisely, Boomi, Komprise and NetApp will sell you excellent instruments for scoring, tracing and serving data to agents, and you should probably buy some of them. Instruments don't decide which workflow goes first, which forty fields form the contract, who owns the vendor master when two agents collide on it, or how narrow the credentials should be when the agent product ships with every connector switched on. Those are architectural decisions about how your company turns its own messy records into decisions a machine can make safely. They have to be designed around your workflows, your systems of record and your tolerance for error, and they compound. The contracts you write this quarter become the semantic spine your company runs on in two years.
Agor AI Advisory does that design work with executive teams. We pick the first workflow with you, write the contracts, scope the agents, and build the shared definition layer as it emerges, so the readiness you build is attached to revenue from the first quarter.
You will never clean the basement. Go and get the stove.
Sources
- Pure AI, "AI Returns Outpace Data Readiness and Oversight, Further Survey Finds," October 19, 2026
- Blocks & Files, "Komprise: AI needs AI-ready data and there isn't enough of it," October 19, 2026
- iTWire, "The Agentic AI Era Is Here, Enterprise Readiness Is Not, New OutSystems Research Finds," October 19, 2026
- Precisely press release via EQS News, "Precisely Introduces New Unified Data Management Platform to Make Enterprise Data AI-Ready," October 6, 2026
- Boomi, "Boomi Acquires Elementary Data to Bring Trusted, Governed Data to the AI Era," October 19, 2026
- TechCrunch, "Google brings agentic AI to Gemini, starting with businesses," October 8, 2026