Bad AI Advice Stops at the Front Line

Bad AI advice stops at the front line, and the research published this autumn suggests that is one of the most valuable things happening in CRM data management right now. On 25 August 2026 Validity published its State of CRM Data Management in 2026 report, a survey of 500 B2B and B2C marketing professionals across the United States, the United Kingdom, Brazil, Australia and New Zealand. The headline most readers took away was that two out of three organisations increased the number of marketing decisions delegated to autonomous AI agents over the past year. The more useful finding sits one level down in the data, and it is genuinely encouraging: the people closest to the records are already catching the recommendations that should not ship.

Validity asked respondents whether they had ever presented or acted on an AI-generated recommendation they later suspected was wrong because of the data underneath it. Sixty-two per cent said yes, 19 per cent frequently and 43 per cent occasionally. Split those answers by seniority and the picture changes completely. Ninety-two per cent of SVPs and VPs said it had happened to them, 78 per cent of the C-suite said the same, and 41 per cent of individual contributors did. Read as a failure, that spread is a governance gap. Read as a capability, it is a working quality control function that nobody has written down, staffed or rewarded, sitting in exactly the part of the organisation that touches the records every day.

What did Validity’s 2026 CRM data report find?

Validity’s State of CRM Data Management in 2026, published on 25 August 2026, surveyed 500 marketing professionals in five countries about the relationship between CRM data quality and AI output. Two out of three organisations had delegated more marketing decisions to autonomous agents over the previous year, while only 21 per cent described their CRM data as very well prepared to support AI. Sixty-two per cent had acted on or presented an AI recommendation they later doubted because of the underlying data, and the rate rose sharply with seniority: 41 per cent of individual contributors, 78 per cent of the C-suite and 92 per cent of SVPs and VPs. Sixty-two per cent also said poor CRM data had probably or definitely cost their organisation revenue.

Two caveats keep the reading honest. These are self-reported suspicions rather than audited errors, so the survey measures how often people doubted an output, not how often an output was actually wrong. And the sample is marketing, so the figures describe campaign decisions more precisely than they describe pipeline or service. Neither caveat weakens the practical point, because a doubt expressed at the right moment is worth more than a post-mortem. Dun and Bradstreet’s AI Momentum survey of 10,000 businesses across 32 countries, published on 28 July 2026, found more than three quarters reporting measurable AI return while just 6 per cent called their data fully ready to scale it. Returns are arriving ahead of readiness, which means the organisations that close the readiness gap get to keep more of the return.

What is CRM data quality, and why does it decide what AI can do?

CRM data quality is the degree to which the records in your CRM are accurate, complete, consistent, unique and current enough for the decisions people and systems make from them. In practice it covers six things: whether a contact exists once rather than four times, whether required fields are populated, whether values follow a single convention across teams and countries, whether the record reflects the customer as they are today, whether relationships between accounts, contacts and opportunities are intact, and whether the reason codes people enter mean the same thing to everyone. It matters more than it used to because an AI recommendation inherits every one of those properties silently. A human reading a stale account record notices the staleness. A model reading the same record returns a confident answer built on it.

That is why data quality has quietly become the highest-leverage CRM investment available this quarter, and why it is a better place to spend than another pilot. Validity found that a third of respondents already spend six or more hours a week fixing data, which is roughly a day a week of skilled attention the organisation is already paying for. Formalising that work does not add cost so much as redirect effort that is currently invisible. It also pays twice, because the same cleanup that makes an agent trustworthy makes forecasting, routing and segmentation better whether or not a single agent ever ships. We made the broader version of that argument in Adoption Is Your Best CRM Investment, and the Validity data strengthens it.

The people nearest the records are already your best control

The 51-point gap between SVPs and VPs on one side and individual contributors on the other is the most actionable number in the report. The straightforward explanation is proximity. A marketing executive who owns the campaign calendar sees a recommendation and a deadline. A specialist who maintains the segment sees a recommendation and recognises the five accounts in it that were merged last quarter. Both are acting rationally on what they can see. Only one of them can see the thing that matters, and in most organisations that person has no sanctioned route to say so before the decision is taken.

Building that route is cheap, fast and almost entirely organisational rather than technical. It takes three decisions. First, name the AI-influenced decisions that are worth a second look, which is usually a short list: audience selection, lead routing, forecast adjustments, pricing suggestions and anything customer-facing at volume. Second, name the person closest to the records behind each one and give them an explicit right to flag, with a service level measured in hours rather than meetings. Third, log the flags. A flag is a free data quality defect report from the person best placed to write it, and after a month the log tells you which fields, objects and integrations are actually generating doubt.

Validity’s respondents are already pointing their budget in a compatible direction. Thirty-nine per cent named real-time monitoring as the capability they most need, rising to 47 per cent of the C-suite on automated monitoring, while 23 per cent favoured consolidating onto a unified platform and 19 per cent preferred third-party validation. Monitoring and human flags are complements rather than alternatives. Automation catches the defects you can define in advance, such as nulls, duplicates and formatting breaks. The person who knows the account catches the ones you cannot define, which is where the expensive mistakes live. Running both is how a team moves from reacting to AI output to governing it, a pattern we looked at from the assurance side in Your AI Reviews Work. Run Them Sooner.

How do you improve CRM adoption rates when AI is in the workflow?

You improve CRM adoption rates by making the system visibly useful to the person entering the data, and an AI feedback loop is one of the strongest versions of that available. The sequence that works: reduce required fields to the ones that drive a decision somebody can point to, show each team the report their own input produces, give the people nearest the records a sanctioned way to correct AI output, and then close the loop by telling them what changed because they flagged it. Adoption rises when effort has a visible consequence. Measure it on completeness of the fields that matter and on time to first useful report, not on logins, and review it at 30, 60 and 90 days after any change.

The AI layer gives this an advantage that earlier adoption programmes never had. For two decades the honest answer to a seller asking why a field mattered was that it fed a report somebody else read. Now the answer is more immediate: that field shapes what the agent recommends to you next week. Salesloft’s 2026 Revenue Benchmark, published on 2 September and based on 500 United States revenue decision makers, found 55.6 per cent saying the loss reasons entered into their CRM were mostly subjective seller guesswork. That is a field whose value changes the moment a model starts learning from it, and sellers understand that argument faster than they ever understood the reporting one.

How do you measure CRM ROI on a data quality programme?

Measure CRM ROI on data quality the way you would measure any operational improvement: pick a small number of before-and-after measures tied to money, and baseline them before you start. Four work well. Track the share of AI-influenced decisions that get flagged and reversed, which should fall over two quarters as defects are fixed at source. Track hours spent manually fixing data, where Validity’s figure of six or more hours a week for a third of respondents gives you a usable starting benchmark. Track forecast accuracy by stage, because duplicate and stale records distort it in a measurable way. Track campaign and sequence rework, given that 67 per cent of Validity’s respondents reported delayed or scrapped campaigns traceable to data problems.

The attractive feature of this particular programme is that the return does not depend on the AI roadmap landing on time. If your agent plans slip a quarter, the deduplication, the field rationalisation and the governance ownership still pay for themselves in cleaner reporting and fewer compliance questions. Validity found 63 per cent citing compliance exposure from poor data and only 41 per cent with a dedicated governance owner or team, which means naming an owner is often the single highest-value hour of work available. That asymmetry, where the investment pays whatever happens next, is what makes data quality the easiest AI business case to defend in a budget meeting.

Why use an independent CRM partner for CRM data work?

An independent CRM partner is a consultancy that implements and advises across more than one CRM platform and does not earn licence revenue from any of them, so its recommendation follows the client’s situation rather than a vendor quota. On data quality work that independence has a specific value. Every major platform now sells a monitoring, enrichment or unification product, and each one is genuinely good at the problems its own architecture creates. The questions that decide the outcome sit between those products: which fields actually drive a decision, which duplicates are duplicates and which are legitimate subsidiaries, which conventions have to hold across two systems after a merger, and who in the business owns each answer.

Those are judgement calls about an organisation rather than configuration calls about a platform, and they travel. A field rationalisation done properly in Dynamics 365 survives a later move to Salesforce. A flag-and-log loop designed with a marketing operations team in HubSpot works the same way in Sales Cloud. The portable part of this work is the larger part, which is why it is worth doing deliberately rather than as a side effect of whichever migration or AI project happens to be funded. It is also the part most likely to be skipped when the only adviser in the room is paid to ship the platform.

The Sirocco perspective

We work across Salesforce, HubSpot and Microsoft Dynamics 365, and the Validity numbers match what we see in client estates. The organisations getting the most from AI recommendations are not the ones with the cleanest data at the outset. They are the ones who decided early who is allowed to say no, and made saying no quick. That decision costs nothing, it needs no procurement cycle, and it generates a defect log that tells you exactly where to spend the cleanup budget instead of guessing.

Our practical advice is to start with one decision rather than a programme. Pick the AI-influenced decision your team is least comfortable with, name the person nearest its records, give them a documented right to flag and a day to use it, and write down what gets flagged for four weeks. The list you end up with is both your data quality roadmap and your AI governance evidence, and you will have built it from judgement your organisation already possesses. That is an unusually good position to be in heading into next year’s planning, and most teams are closer to it than they think.

Get in Touch

If you are working out how to route AI recommendations past the people who actually know your CRM records, we would be glad to talk it through on Salesforce, HubSpot or Dynamics 365.

So where do you start?

As your long-term partner for sustainable success, Sirocco is here to help you achieve your business goals. Contact us today to discuss your specific needs and book a free consultation or workshop to get started!