Book a discovery call if you need to ship AI-native: meetings-na2.hubspot.com/mar…

We used AI agents to build three systems for a B2B lender with two engineers, inside a platform already running in production. Development cycle time dropped by 50% with each successive build, with no disruption to live operations. The scope covered ACH funds movement, an underwriting portal, and a syndicator portal. A regression could affect live disbursements and collections, so the agents needed a clear understanding of the codebase and a reliable way to check their work. Before feature development, we ran an AI adoption audit on the repository. It found four gaps: missing architectural context, build and test commands the agents couldn’t find, inconsistent patterns across modules, and recurring procedures that hadn’t been documented for reuse. Those findings shaped the harness we built around the agents. Repository and module context files documented the architecture, conventions, and exact commands. A skills library gave agents reusable instructions for recurring tasks. Automated checks and review were built into the delivery process. From there, every change followed the same path: → A written spec defined the work and served as the source of truth for engineers and agents. → Tests came before implementation, making the expected behavior explicit and testable. → Coding agents implemented the feature inside a deterministic verification harness. → Tests, static analysis, type checking, and custom validators ran before the change reached review. → A PR agent and CodeRabbit reviewed every change against the spec and engineering standards. Zero changes merged without automatic verification. All three systems also shared one architecture. Common components reduced duplication, and each successive build started on a foundation that was already in place. We reused the process for directing and checking the agents alongside the code. What reached the business: → Automated ACH processing replaced manual disbursement and collection workflows. → Underwriting moved into one environment for applicant data, risk rules, and team workflows. → Syndicators gained real-time, self-service access to positions, deal participation, and returns. The delivery gains came from that combination: agents doing the implementation work, a harness checking their output, and a shared foundation that reduced how much each new build needed. That’s what our AI Velocity Pods bring into an existing codebase. Our engineers set up the context, specifications, and verification the agents need, then use that foundation to deliver the next capability your business needs.

The 8 stages of every company's first AI agent: 1. "Someone on the team built it in an afternoon." 2. The CEO demos it at the board meeting. 3. It goes live on real customer data. 4. Week four. It starts giving different answers to the same question. 5. Someone patches the instructions. Now it's wrong somewhere else. 6. The person who built it goes on vacation. It stops. 7. The operating partner asks what it saves per task. Nobody can say. 8. "Turns out you can't run a business on a demo."

A US healthtech startup only gets paid when a patient's balance reaches zero and the payment reconciles against the right charge in the provider's billing system. We built the system that does that work: AI agents that call, text, email and mail patients, explain the bill, and set up payment. No fee for messages sent or calls made. The money comes when the books close. They talk to patients on the provider's behalf, under a BAA, about bills for care they already got. Get it wrong and it costs the provider a patient. Discovery started in August 2025. On December 4 the first message went to a real patient. Three days later the voice agent made its first call. It hasn't been switched off since. I know most people don't trust an AI agent on the phone with patients about money. So here's what sits under it. Six layers, and the agent is only the top one: 1. The voice agent. Outbound, inbound and a scheduling line, with 12 tools it can use mid-call: find the account, pull charges and insurance adjustments, text or email a payment link, build a payment plan, send a statement, book an appointment. 2. Shared context. Every channel reads the same case: the visit, what insurance paid and why, every payment, every past call and text. A patient who calls back never explains twice. 3. Billing system integrations. Modern APIs, FHIR, SOAP and nightly file exports, all behind one interface. Bad rows get quarantined, never silently dropped. 4. Money. Exact decimals and double-entry ledgers. No floating-point math touches money, anywhere. 5. Reconciliation. Drift checks compare the payment processor, the ledger and the billing system. A repair job refuses to run unless the books agree first. 6. Audit and consent. Quiet hours enforced at send time, consent tracked per channel, and 6.36 million audit records with no patient data in the logs. The safeguards live in code, where a prompt can't talk its way around them. The agent says nothing about a balance until the caller's identity is verified. A retried request can't run twice. Test agents can't write to a real patient record. Disputes, lawyers and requests for a person go straight to staff. AI let our engineers write code faster than anyone could check it. So we grew the tests until they were almost as big as the app. 349,042 lines of application code. 343,080 lines of test code. 13,650 automated tests run on every pull request in 9.7 minutes, and every one of the 216 database migrations is proven safe to run twice. The system also reviews itself. Models read 5,135 conversation threads and logged 835 findings. A person approves every fix, and a separate model checks fresh calls afterward to confirm it worked. By late September: → 61,246 AI voice calls → 658,657 of 949,563 scheduled messages and calls cancelled before sending, because the patient paid, disputed or the account changed → Around 97% of patient conversations finished without a human → 18,194 balances driven to zero If you don't trust an AI agent on the phone with your patients, you shouldn't. Not until everything above is in place.

AI agents can unlock serious value in business operations: faster collections, fewer billing errors, shorter close cycles, and more orders handled by the same team. If I were looking for where to start, these workflows would be high on the list: > accounts payable: read incoming invoices, match them against purchase orders and receipts, and flag discrepancies for review before payment > cash application: gather payment details from emails and documents, match payments to invoices, and investigate amounts that don't line up > collections: prioritize overdue accounts, check recent payments and open disputes, and follow up with the relevant account history > billing disputes: gather the invoice, customer correspondence, and supporting records, then recommend a resolution for the team to approve > revenue leakage: compare draft bills against contracted rates and statements of work, flagging missing charges and pricing discrepancies before invoices go out > account reconciliation: identify differences between subledgers and the general ledger, investigate exceptions, and recommend corrections for review > month-end close: trace unusual balances back to the underlying transactions and prepare explanations for the finance team > expense processing: read receipts, match them to card transactions, check company policy, and route exceptions for review > purchase requests: collect missing details from employees, guide them toward approved buying options, and route requests through the right approval process > supplier follow-ups: chase unconfirmed purchase orders, read supplier replies, and propose updates to delivery dates and quantities in the ERP > sales order entry: turn customer emails into quotes and orders, check availability, and ask for missing details before processing > order exceptions: investigate failed or delayed orders, identify the cause, and recommend or execute corrective actions within agreed permissions > IT service requests: gather the information needed to resolve routine issues, run approved fixes, and hand unresolved cases to the right team with context What makes these workflows worth looking at is how much time your team spends gathering information, checking it against another system, and chasing someone before a case can move forward. At @LimestoneHQ, we build agents around those workflows, connecting your existing systems and keeping your approval rules in place. We work month to month, with the baseline, measurement period, and guarantee agreed in writing: no measured uplift, no invoice.