Vellm.ai CEO David Flickinger says a roofing AI estimator produced a $6.5m bid in 23 minutes, within $400 of the two-week human estimate.
The roofing company doubled its commercial win rate and expects to double next year without hiring.
Flickinger says a memory layer of owner decision logic can help buyers value SMBs more fairly during diligence.
AI pays off when it is wired into the business
Vellm.ai sells managed AI infrastructure to small and mid-sized businesses, and its founder's case is that value sits in connection, not conversation. An agent wired into a company's CRM and financial documents priced a $6.5m roofing bid in 23 minutes and landed within $400 of a two-week human estimate, drawing on a memory of 3,700 past proposals. The same architecture captures an owner's decision logic in a form a buyer can query during diligence, which should narrow the discount buyers apply to businesses they do not understand. Cost is the constraint.
Takeaways
Connected AI beats prompting alone. Mikk found 11 of 15 Northern European owners using AI, mostly basic ChatGPT. Flickinger argues the real value comes when agents connect to the CRM, ERP, databases and financial documents so answers carry the business's own context. Prompting alone produces confident wrong answers.
Roofing estimate took 23 minutes. On a South Florida job of 47 buildings, the human estimator spent about two weeks. The AI produced a $6.5m number in 23 minutes and landed within $400. The roofing company has doubled its win rate and expects to double next year without hiring.
Human checkpoints stay until validated. Vellm builds review steps into workflows and removes them as they are validated, and no client yet lets AI email proposals unreviewed. Junior estimators spend more time with customers while the chief estimator reviews a proposal in a minute or two.
AI memory layer can raise valuations. Buyers discount businesses partly because they do not understand what drives risk and success in them. A well-deployed architecture distils the owner's decision logic into a memory layer a buyer can query during diligence.
Cost is the first AI risk. Flickinger puts cost ahead of data security. Frontier labs are soon-to-be-public companies with shareholders, which pushes high per-token fees, and one document workflow went from a few cents to about $14 per iteration. Routing tasks to the right model, including open-source ones, controls spend.
Connect agents to systems before automating. Owners should first connect an agent to their systems, then automate work streams as agent loops. Off-the-shelf tools and subscription AI create no advantage because competitors run the same ones.
How the roofing estimator works
An orchestrator agent owns the bid and tasks sub-agents with each piece of the estimate.
A plan reader extracts numbers from the architectural plans of a commercial construction project.
One agent references similar past projects. Others pull current materials and labour pricing.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
A risk agent weighs travel distance and crane needs.
The memory bank was built from more than 3,700 proposals, with rules extracted from spreadsheets and corrections learned from job costing, such as a job estimated at a 16% margin that came in at 12% because labour was underestimated.
What Flickinger says to do first
Connect an agent into the business's own systems before anything else, so answers carry company context.
Automate whole work streams as agent loops rather than buying off-the-shelf tools everyone else can buy.
Build review checkpoints into each workflow and remove them only once they have proven out.
Route each task to the right model, mixing frontier, open-source and local options to control cost.