When an AI agent gets something wrong, the first instinct is to blame the AI. Wrong model, wrong prompt, wrong vendor. In almost every failed agent we've been asked to look at, the model was fine. The agent was handed three versions of the same customer, a deal stage nobody had updated since winter and a notes field full of half-sentences, and it did exactly what you'd expect. It got confused.
There's no shortage of evidence that AI projects are hard to land. Gartner predicted in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027. McKinsey's latest State of AI survey found 88% of organisations now use AI, but only 39% report any impact on EBIT. And the MIT NANDA research reported by Fortune found that 95% of generative AI pilots saw no P&L return, with the authors themselves noting the sample was small and the finding directional.
Read those together and a pattern shows up. Using AI is easy. Getting a return from it is not. The gap between the two is mostly about what the AI is given to work with.
The models available today are remarkably capable. They can read an email, summarise a phone call, draft a reply and decide which of five options fits best. What they can't do is know things that aren't written down, or tell which of two conflicting records is true.
An agent's job is to take a goal ("reply to this enquiry", "chase this overdue invoice") and work out the steps using the information in front of it. If that information is wrong, incomplete or scattered, a better model just produces a more confident wrong answer. Swapping models is like replacing a new staff member because they couldn't find a file that was never filed.
These are the ones we see most often, in businesses of every size.
The same customer exists three times: once from a web form, once from an import, once typed in by hand with a different email. The agent picks one, and it's often the wrong one. It tells a loyal client they're a new lead, or misses the open job sitting on the other record.
The deal exists. The contact exists. The invoice exists. But nothing joins them together. A person can piece it together from memory. An agent can't. In HubSpot, associations are the lines between records that tell an agent "this invoice belongs to this job, for this company, managed by this person". Without them, every record is an island.
Deals that closed months ago still sit at "Proposal sent". Tickets that were resolved by phone are still open. An agent working the pipeline will chase people who already said yes, or no. Nothing damages trust faster.
"Client wants start mid-March, budget approx 40k, talk to Sam not Jo." That's useful, but it's buried in a notes field. Important facts that live only in notes are facts an agent might miss or misread. Key details belong in proper properties with clear names and set options.
If your own team can't agree when an enquiry counts as qualified or a job counts as complete, an agent has no chance. This one isn't really a data problem, but it shows up in the data as inconsistent stages and half-filled fields.
This is the core of how we work. We build agents on top of HubSpot because it gives an agent a structured map instead of a pile of paper. Sales pipelines, service pipelines, project pipelines, tickets, invoices and custom objects all live in one place, joined by associations. An agent can see what kind of record it's looking at, follow the links to the related records it needs, and know exactly where each piece of work is up to.
That structure also makes it possible for agents to hand work to each other. A sales agent can close a deal and pass a clean, fully associated record to a service agent to start delivery, the same way a good salesperson briefs the delivery team. When the data is messy, every hand-off loses something.
None of this is new thinking. We wrote about data management in the age of AI a year ago, and about how sync properties keep associated records consistent before that. What's changed is the cost of ignoring it. A messy CRM used to mean bad reports. Now it means an agent making bad decisions on your behalf.
You don't need a perfect CRM. You need a clean one for the job the agent is doing. Start narrow:
This is the Context step of our method: check the data, pick one job, agree what done means and measure today's cost. Then comes Action (HubSpot's own agents first, tested on real past cases, starting in draft), Coordination (people approve, agents act) and Result (a monthly scorecard showing what the agent did, saved and cost). Skip Context and the rest doesn't matter.
If you're new to agents altogether, our plain-English guide to AI agents covers the basics. And for a sense of how data quality plays out in one industry, see membership retention agents, where the agent is only as good as the attendance and billing data behind it.
If your agent is underperforming, look at the data before you look at the model. More often than not, the fix is a merge, an association or a cleaned-up pipeline, not a new vendor. And if you haven't built one yet, the best investment you can make is getting the CRM right first. It pays off whether you end up using agents or not.
Want to know if your data is ready? Our AI Readiness Audit checks your HubSpot for duplicates, missing associations and stale pipelines, then tells you which job an agent could take on first. If you're already confident in your data, go straight to Build my solution, or see how we work.