Build, license or wait on AI dispatch
The decision
Should we build our own AI dispatch agent, license a white-label dispatch engine, or hold AI dispatch until after our Series B, given that we need it live with paying customers by April 2027?
License Vendor A now on short, non-exclusive, no-training terms to hit April 2027. Keep your data and ranking talent in-house to build the eval harness and future replacement. Refuse Vendor B and do not hold.
Why
The real tension is Expansionist ownership against everyone else's calendar.
The calendar wins on your own numbers: 54 engineer-months with 4 free engineers, no LLM shipping history and 4-month ML hires lands after the board's April 2027 bar. Holding fails a board that has rejected a demo, and Vendor B puts a legal amendment and a roughly 40% opt-in on the critical path while handing your one moat to a vendor that can serve your rivals.
The Expansionist is right that the data is the asset, and keeping the eval harness and a future ranker in-house protects it, but the bridge has to be the plan, not the fallback. Several numbers in the council are hypotheses, not facts: the Expansionist's claim that 25% acceptance is 'easy', the Buyer's $5-10 per tech willingness to pay, the Researcher's 23% adoption figure, and both average-technician counts.
The council found no benchmark for 25% acceptance, so treat that bar as the unproven part of the plan.
Biggest risk
Adoption, not shipping. A generic engine that doesn't understand HVAC certifications, job duration or callback risk produces suggestions dispatchers stop accepting within weeks, so you reach 20 live customers but miss 25% accepted dispatches, or reach it only through UI defaults that diligence will see through. Secondary risks are vendor lock-in and the vendor also serving your two competitors.
Biggest upside
Real usage before the May 2027 raise, plus a credible ownership story: a licensed engine proves demand now while your four years of labeled outcomes (assignments, drive times, completions, callbacks) train an in-house ranker you swap in after the B. Investors get both traction and a moat. Paid add-on attach later could lift NRR from 104%.
The cheapest test of the riskiest assumption
In 48 hours, have your two route-optimisation engineers replay 3 months of history for 5-10 representative customers.
Score a naive baseline (closest qualified, available tech) on how often the dispatcher's actual pick was the top-1 suggestion and how often it was in the top 3, segmented by job type. If top-1 agreement is well below 25%, the acceptance bar is the real risk and any vendor must clearly beat this baseline.
In parallel, send Vendor A the same replay set under NDA with a no-training clause, after a quick counsel check that eval-only processing is permitted, and ask for their top-1 agreement plus drive-time and callback deltas before you sign. Also call 10 candidate design partners (20+ techs, busy dispatch desk) with a 90-days-free offer and count the yeses.
The money read
Your ~$8k ARR per customer implies you charge roughly $10-20 per tech per month today, so Vendor A's $4-6 per tech is material and must be priced as a paid add-on, not bundled free.
Bundled across the base it could cost $1-2M a year and erode the 78% margin. Pilot cost is trivial: 20 customers at about 30 techs is roughly $30-40k a year.
At add-on prices under $25 per tech, the 20% revenue share costs less than the flat fee, so negotiate revenue share or a free pilot through the raise, with an option to convert to flat per-tech pricing. Timeline: sign by mid-November, shadow mode with design partners in December, live and approving in January, giving about 3 months of usage by April.
First AI dollars arrive around Q2 2027 if you run 90 days free, which is fine because the board doesn't require revenue. Shippable with the 4 free engineers on integration, the approval UI and instrumentation; do not touch the mobile rebuild team.
The council
Five AI roles argued the decision; an AI judge then had to rule GO, RESHAPE or KILL. “It depends” is not an allowed answer.
The Contrarian
Assumes it failsBuild and hold both fail the April 2027 date on the arithmetic, Vendor B is a legal and diligence trap, and Vendor A is the only option left, so it gets picked by default and then fails on adoption.
What they insist you hearOnly licensing can hit April, and only if you sign by about November with a pilot live by January. Start by running Vendor A's engine offline against your four years of history to see if it beats dispatcher decisions on drive time and callbacks. If you can't show that before signing, the usage number will be hollow and diligence will catch it.
The full argument · 5 points
- Build misses the date. Nine months from today is July 2027, and that assumes 6 engineers. You can free 4, a fifth and sixth would break the Q2 mobile promise, and ML hires take 4 months to land. Nobody on the team has shipped an LLM product. The board wants 20 customers live at a 25% accepted-suggestion rate by April. That means shipping by about February to leave weeks of usage data, so you would be 5 months late with a shortfall of engineers.
- Vendor B's price depends on rights you don't have. Your contract covers using de-identified data for your own service, not for a third party's model training. Counsel says it needs an amendment plus opt-in. If 40% opt in within 6 months, that is a self-selected sample that probably skews toward your most engaged customers. Asking 1,100 contractors for training rights will also raise privacy questions and could hurt NRR (104%), which is already mediocre. Series B diligence will ask why you gave your four-year dataset, your main moat, to a vendor that can sell the resulting model to your two competitors.
- Vendor A makes the usage real but rented. Take 20% rev share on a product you resell and the 78% gross margin on that line drops sharply. At $4-6 per technician per month across a base of roughly 30 technicians per customer, the cost is about $2M a year at full rollout. Investors will ask what you own. Lock-in, data ownership, exit terms and whether the vendor also serves your competitors are all unresolved. A generic engine that doesn't know HVAC callbacks or job-type duration may not beat your existing rules-based scheduling.
- Holding until after the Series B guarantees the worst story. You would walk into a May raise with no AI usage while two direct competitors shipped in 2026, and the board has said a demo doesn't count. It also leaves your 1,100 customers to be approached by competitors with AI dispatch for another 6+ months.
- The 25% acceptance metric is the real failure point, and none of the three options addresses it. Dispatchers override suggestions out of habit and distrust, and in recommend-only mode every override counts against you. Twenty pilot customers will be cherry-picked, and a number built that way will not survive diligence. The model needs to be demonstrably better than a dispatcher's own judgment on drive time and callbacks, and nobody has tested that on your four years of data.
The Expansionist
Makes the strongest case for itBuild the recommend-only dispatch agent in-house, and use a licensed engine as a fast bridge. Your four years of job data is the moat, and the way to monetize it is to own the agent and sell it as a premium add-on, not rent it.
What they insist you hearThe thing you must protect is your data and your ownership of the dispatch decision. Treat April 2027 as a test of whether you can move fast on that, not as the goal itself. Build the ranking core with the engineers you can free, rent Vendor A only as a bridge, and refuse Vendor B. Waiting for the Series B means pitching investors a story your two competitors have already shipped.
The full argument · 5 points
- Your data is the asset. Four years of assignments, drive times, completion times and callbacks across 1,100 contractors is a labeled outcome dataset. Competitors' models trained on generic data can't match it. Vendor B wants this data for free, and Vendor A's engine would be generic until it is tuned on it. Selling your data cheaply, or building on someone else's engine, hands away the one thing that makes AI dispatch defensible.
- The money is in the add-on. Assume about 15,000 technicians at 1,100 customers averaging about 14 techs. If you charge $15-25 per tech per month for AI dispatch and reach 30% attach, that is roughly $0.8-1.4M of new ARR on a $9M base. That would lift NRR from 104% to 110% or more, which is what Series B investors pay multiples for. Vendor A's $4-6 per tech cost against that price keeps a healthy margin, and a 20% rev share caps your margin and your upside.
- The 10x version is the system of record becoming an autonomous operations layer. Recommend-only comes first. Then come auto-assign for routine jobs, then dynamic re-dispatch when a job runs long or a tech calls out, then callback-risk prediction, parts-on-truck matching, and pricing and upsell signals. Each step uses the callback and completion data you already hold. Autonomy is also what justifies a usage-based or outcome-based price. Hold this for after the raise and show it as the roadmap.
- Hit the board's bar with a design-partner program starting now. You need 20 live customers with 25% or more accepted suggestions by April 2027. Pick 30 customers in the 30-100 tech range, offer them free or discounted access in exchange for usage commitments, and ship a shadow-mode version in about 4 months. Shadow mode is the agent suggesting without dispatchers acting on it. It builds trust and gives you acceptance-rate data. The acceptance metric is easy to hit because dispatchers already pick from a short list of techs.
- Use a hybrid build. The 4 free engineers and the 2 who have shipped route-ranking features can build the core, and the problem is mostly a ranking and optimization problem, not an LLM problem. Hire 1-2 ML contractors now. License Vendor A on a short, non-exclusive pilot as a fallback or benchmark for a handful of customers. That protects the April date, and you can swap it out once your own model beats it. Don't touch Vendor B. It needs contract amendments, only about 40% of customers would opt in, and it creates trust and privacy risk with contractors who see you as their system of record.
The Logician
Reasons from first principlesLicense Vendor A on a flat per-technician fee now; building misses April by arithmetic, holding fails the board's own condition, and Vendor B trades your only moat for a contract you can't legally sign yet.
What they insist you hearOnly one option fits the calendar: licensing. You have about 6 months, and the usage has to accumulate before the raise, so sign by November, ship to a pilot cohort by January, and negotiate exit and data-portability terms now. Meanwhile, keep the dataset in-house and use your ranking engineers to build the evaluation harness and a future in-house ranker.
The full argument · 5 points
- Build math fails: 6 engineers x 9 months = 54 engineer-months. You can free only 4, so that is about 13.5 months, and hiring ML engineers takes 4 more months. Even at full staffing, 9 months from today is July 2027, after the April deadline. The deadline is a hard constraint and this option violates it.
- Holding fails the objective: the board wants 20 paying customers live with at least 25% of dispatches accepted via suggestion, and usage must accumulate before May 2027. Waiting until after the Series B means no usage to show at the raise, and two competitors already shipped. Waiting only helps if the raise doesn't depend on AI, and your board says it does.
- Vendor B is logically broken for the timeline: training on customer data needs a contract amendment plus opt-in, and only about 40% opt in within six months. That is a legal dependency on the critical path. It also gives a vendor, who can sell to your competitors, the four-year dataset that is your one real advantage. Cheap per-seat pricing is paid for with your moat.
- Vendor A economics: the flat fee and the 20% rev share cross at a price of about $25 per tech per month (20% x $25 = $5). Your implied revenue is about $8k per customer per year, so you likely charge well under $25 per tech. Rev share therefore costs less than the flat fee at low add-on prices, and the flat fee wins only if you can price the add-on above ~$25. Twenty customers of about 40 techs each is roughly $50k a year, which is trivial. Full rollout at 1,100 customers could cost $1.5-2.5M a year, which threatens the 78% margin if you don't charge for it.
- The 25% accepted-suggestion metric is gameable and depends on suggestion quality plus UI defaults. Recommend-only dispatch is a ranking problem (assignments, drive time, callbacks), not an LLM problem, and your two route-optimisation engineers are suited to it. Use them to build an evaluation harness on your four years of history and benchmark the vendor against it. Track callback and drive-time lift as well, so the usage is real and not just clicks.
The Researcher
Works from web evidenceThe real world says license now. Competitors already ship AI dispatch, and a 9-month build cannot meet an April 2027 deadline that is 6 months away. Holding means no usage story at all.
What they insist you hearDon't build the whole thing, and don't give away your data. Sign Vendor A, or an equivalent engine, on per-tech pricing by about November, with a no-training clause and a data-portability clause. Put your 4 engineers on integration, the dispatcher-approval UI and instrumentation of the accepted-suggestion rate. Use your four years of job history to build your own ranking and evaluation layer for a later swap. If Vendor A's contract can't be signed and live in pilot by about January 2027, the 20-customer bar is already at risk.
The full argument · 5 points
- Competitors: ServiceTitan sells Dispatch Pro, an AI dispatch add-on that re-evaluates assignments through the day using skills, location and job value (https://www.retellai.com/blog/hvac-dispatch-software, https://www.servicetitan.com/features/pro/dispatch). One roundup says Jobber, Housecall Pro and Workiz bundle or sell AI voice agents for call answering, which is not dispatch (same Retell URL). One vendor roundup says every major FSM platform now ships similar AI features, so 'we have AI dispatch' is no longer a differentiator (https://gigacatalyst.com/blog/field-service-ai-features-2026). That supports building something now, but the value comes from usage and outcomes, not from having the feature.
- Demand is real but not proven for your segment. A vendor blog (exoserva.com) claims AI scheduling adoption among contractors rose from 8% in 2023 to 23% (https://exoserva.com/blog/state-of-field-service-2026). That is a vendor-published figure, so treat it as UNVERIFIED. The same page cites a $5.2B FSM market in 2025. Mordor puts it at $5.66B in 2025 (https://www.mordorintelligence.com/industry-reports/field-service-management-market). ServiceTitan user feedback on Dispatch Pro is mixed. One Reddit user says you still need a dispatcher but can run roughly 30+ techs per dispatcher (https://www.reddit.com/r/HVAC/comments/1jp97gv/service_titan_with_ai_dispatching/). A Facebook group snippet says it works with thorough setup and 20+ techs, while others report problems (https://www.facebook.com/groups/455916369530577/posts/1569840971471439/). Expect adoption to need onboarding effort, which puts the 25% accepted-suggestion bar at risk. I found no public benchmark for a 25% acceptance rate, so that bar is UNVERIFIED as realistic.
- Build fails on your own numbers. It is Oct 5, 2026 and the board bar is 20 live customers by April 2027, about 6 months away. Six engineers for 9 months ends around July 2027 even at full staffing. You can only free 4 without delaying the Q2 mobile rebuild, and nobody has shipped an LLM product. Hiring ML engineers takes about 4 months. That is my arithmetic from your inputs, not a sourced figure. A build also leaves no time for live usage before the May raise.
- Licensing is cheap relative to the stakes. Vendor A at $4–6 per tech per month is far below the roughly $100–150 per tech per month that one source reports for FSM software, and $245 per tech per month list for ServiceTitan (https://cal.com/fr/blog/best-field-service-scheduling-software, https://www.lvgps.com/blogs/news/field-service-dispatch-software). Both are third-party estimates. My illustration for 20 customers at an assumed 25 techs each (UNVERIFIED assumption): 500 techs × $5 × 12 ≈ $30K a year. Per-tech pricing beats the 20% revenue share once you price an add-on. Embeddable engines exist as an alternative or second source. Timefold markets an FSM routing API and claims about 25% less travel and a 2–4 week integration (https://timefold.ai/blog/field-service-routing-api-the-scheduling-engine-behind-high-performance-field-service-management-platforms). Those are vendor claims, and I found no public Timefold pricing.
- Vendor B is the wrong trade. Your counsel says it needs a contract amendment and customer opt-in, and you estimate only about 40% would opt in within six months. That is the same window as your deadline. You would also be giving a vendor your four years of job history, which is your one durable asset and a core Series B story. Timefold's own writing argues that scheduling needs optimization and not just LLMs. That fits the ranking-feature experience your two engineers already have, so keep that skill in-house.
The Buyer
Speaks as the counterpartyAs a 40-tech HVAC owner, I'd try a recommend-only dispatch assistant from my current software vendor. I won't pay real money until it has shown me it beats my best dispatcher, and I don't care whether you built it or licensed it.
What they insist you hearThe customer won't pay for 'AI dispatch'. They'll pay when a dispatcher accepts suggestions because they're explained and usually right. Pick whichever route gets a trustworthy recommend-only product to about 20 design-partner shops by early 2027 without using customer data you haven't earned the right to share. Licensing Vendor A while using your own four years of data to tune and validate it looks safest to me. Skip Vendor B's training clause: the amendment and opt-in process would slow you down, and its data terms would cost you the trust you need.
The full argument · 5 points
- My real objection is trust, not features. Dispatch is the job that keeps trucks rolling and customers happy. If the tool sends my best HVAC tech to a drain clear, my dispatcher stops clicking accept within a week. I'd want a one-line reason on every suggestion (closest, certified for this equipment, lowest callback rate) and an override that takes one click. Without that I'd do nothing and keep my dispatchers.
- I'd pay for measurable savings, not for 'AI'. A 30-tech shop pays roughly $150-200k a year for dispatch labor and drive time (a guess). If you show me fewer windshield hours, more jobs per tech per day and fewer callbacks, I'd pay about $5-10 per tech per month as an add-on (my guess, not a benchmark). That's close to Vendor A's quoted $4-6 cost. A 20% revenue share on top means you'd need to charge more than I'd accept unless the savings are proven. Free for 90 days, then paid, gets a yes from me today.
- Vendor B's data-training clause would make me hesitate. My job history, customer addresses and pricing patterns are my business, and 'anonymized' doesn't reassure me when a rival could end up using a model trained on my data. I'd probably opt out unless you named the vendor, told me what's shared and gave me a discount or a free tier. Your own 40% opt-in estimate suggests many owners feel the same. Whatever you pick, my four years of history should be improving the model you give me, and that part your current contract already covers.
- What would send me to a competitor is that two rivals already ship AI dispatch. From memory, and unverified, ServiceTitan has a dispatch-automation product. If I'm weighing a switch and a rival's AI is live while yours is 'coming soon', that tilts me. If I'm happy and my crew is used to your software, I'll wait for you, but not past mid-2027, because my mobile app and my dispatcher's habits are the real switching costs. Don't delay the mobile-app rebuild you promised me for Q2 2027 just to ship AI. I asked for the app first.
- The 25% acceptance bar is achievable only with design-partner customers. Getting 20 of 1,100 shops live is easy if you ask nicely and discount. Getting dispatchers to accept a quarter of dispatches means the suggestions have to be good from day one, with a dispatcher in the loop who gives feedback. Pick customers with 20+ techs, a busy dispatch desk and clean data, and onboard them personally.
Method and verification
Council
Five AI roles argue the decision, then an AI judge rules GO, RESHAPE or KILL and names the cheapest test of the riskiest assumption.
What happens next
Follow-up questions
One round of follow-up questions is included, answered in writing, sent within 14 days of delivery. Email them together to info@ckkcassociates.com with your project reference.
Prepared with AI research methods and reviewed by a person before release. Citation and claim verdicts are model assessments; they do not establish that every statement here is true. Grades describe what backs a finding, not its probability.