Lightfield in Inc.: Why AI Agents Fail on CRM Data Built for Humans

Inc. published Your CRM Was Built for a World Before AI. That's Why Your Agents Keep Failing, a feature by Hillary Remy on why AI agents deployed on top of a CRM so often disappoint. Lightfield's co-founder and CEO, Keith Peiris, is the expert the piece turns to when it asks what agents actually need. AuthorityTech is Lightfield's earned-media partner and secured this placement.

The question a revenue leader is actually asking

A team rolls out an AI agent for pipeline review, account research or follow-up. The output is off: stale context, wrong account details, deal stages that do not match what the reps know. The natural move is to fix the AI by rewriting prompts, reconfiguring the agent or trying a different vendor.

The Inc. feature argues that this is usually the wrong place to look. Its thesis is that the failure sits underneath the agent, in the CRM, and that most teams "don't figure this out until they've spent several months debugging the wrong thing."

That matters to anyone deciding whether to keep investing in agents on their current CRM or to change the system the agents read from.

What the Inc. feature says

The piece opens with Gartner's June 2025 projection that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. It then locates much of the trouble in CRM records themselves:

• Close dates that have stopped meaning anything because they have been pushed back so many times.
• Notes written as reminders for whoever took the call, "not as a record another system could make sense of."
• Fields that exist in the interface but are blank in the database.

Its explanation for why this was tolerable before agents is the most useful part. CRM systems were built on the assumption that salespeople would reliably enter their own data, and they do not. Humans reading a CRM compensate: a rep picking up a deal mid-cycle can ask questions, infer from context and guess what a six-week-old note meant. An agent cannot. "A missing field isn't something it navigates around. It's a break in the chain." The gaps humans filled without thinking "become the exact places agents fail."

The feature backs the data-quality point with Salesforce's seventh State of Sales report, a survey of 4,050 sales professionals conducted in August and September 2025. Sales teams using AI agents ranked manual errors as their top data problem, with incomplete data also in the top five, and reps reported spending 60 percent of their workweek on work other than selling, including manual data entry.

Where Lightfield comes in

Under the heading "What agents actually need to work from," Inc. introduces Peiris as co-founder and CEO of Lightfield, "which raised $47 million in a Series A led by Andreessen Horowitz to rearchitect CRM for the agent era." His diagnosis, given to Remy in an interview, is the line the article is built around:

"Agents don't fail because the models aren't capable. They fail because the data they work with is incomplete, inaccurate, and missing the structure needed for comprehension."

That is a specific claim about where the problem lives, and it is the claim Lightfield's product is built on.

What Lightfield does

Lightfield describes itself as an AI-native CRM. On its own site, lightfield.app, the company says Lightfield "updates itself from every customer interaction, so its agents can run outbound, flag deals at risk, and tell you where to focus next." The company's site confirms the Series A as a $47M round led by a16z, and Andreessen Horowitz announced it as leading Lightfield's $47M Series A.

Fact	Where it comes from
Keith Peiris is Lightfield's co-founder and CEO	The Inc. feature
$47M Series A led by Andreessen Horowitz	The Inc. feature, Lightfield's own Series A announcement and Andreessen Horowitz's investment announcement
The CRM updates itself from every customer interaction	Lightfield's homepage, lightfield.app
Its agents run outbound, flag at-risk deals and recommend where to focus	Lightfield's homepage, lightfield.app

The connection between the article and the product is direct. If the reason agents fail is that reps never entered complete, structured records, then a CRM that captures the record from the interaction itself, rather than from what someone chose to type, removes the dependency the feature identifies. Whether it does that well enough for a given team is something a buyer tests, not something an article settles.

Why an Inc. feature reaches Lightfield's buyer

Inc. is read by founders and operators running growing companies, which is close to the buyer who chooses or replaces a CRM and then has to live with whether its agents work. A DA-92 domain also tends to rank for the question in the headline and to be drawn on by AI assistants when someone asks why their sales agents keep getting things wrong. For Lightfield, that puts its CEO's explanation of the problem in front of buyers at the moment they are diagnosing it.

What to check before blaming the model

The Inc. feature's argument turns into a practical test. Before swapping an agent vendor or rewriting prompts, a revenue leader can check whether the CRM underneath can support automated reasoning at all.

Check	What to look at	Why it matters for agents
Field completeness	Pull 50 open opportunities and count the blank fields that the agent's task depends on	A blank field is a break in the chain for an agent, not a gap it can reason around
Close-date integrity	How many open deals have had their close date moved more than twice, with no recorded reason	Repeatedly moved dates with no explanation give an agent nothing it can act on
Notes a machine can read	Read ten recent call notes as if you had never met the customer	Notes written as personal reminders carry context only their author can decode
Source of the record	What share of activity is captured automatically from email, calls and meetings versus typed by reps	The feature's core point: the record reflects what someone chose to type, not what happened
Agent-on-data test	Run the same agent task on a clean, fully populated account and on a typical one	If the clean account works and the typical one fails, the model is not the bottleneck
Data-entry load	How much of a rep's week goes to logging rather than selling	Salesforce's survey puts work other than selling at 60 percent of the week; a system that needs that input will keep missing it

If the checks show the problem is in the data, a better prompt will not fix it. The decision then is whether to clean and maintain the existing records by hand or to move to a system that captures them without depending on reps, which is the category Lightfield is building in.

FAQ

What does the Inc. feature say about Lightfield? It names Keith Peiris as co-founder and CEO of Lightfield, notes the company raised a $47 million Series A led by Andreessen Horowitz "to rearchitect CRM for the agent era," and quotes him on why agent deployments stall: the data they work with is incomplete, inaccurate and lacks the structure needed for comprehension.

Why do AI agents fail on CRM data? According to the feature, because CRM records were built on the assumption that reps would enter complete data, and they do not. Humans fill the gaps from context. Agents read what is in the record and act on it, so blank fields, unexplained date changes and private shorthand become the points where they fail.

What does Lightfield do? Lightfield is an AI-native CRM that, in its own words, updates itself from every customer interaction so its agents can run outbound, flag deals at risk and recommend where to focus next.

Is the problem the AI model or the CRM? The Inc. feature argues it is usually the CRM. A simple test is to run the same agent task on a fully populated account and on a typical one. If only the clean account works, the model is not the constraint.