AI agents and systems
AI that still works after the demo
Most of what we’re asked to build starts with a chat window or a phone line, and plenty of it should stay there. Behind that is the bigger job: agents that carry work across your systems, answer from your own records, and can be shown, in numbers, to do what they’re supposed to. We build both, and we hold both to the same standard.
Two kinds of work
From the first call to the work behind it
The first line
Agents your customers talk to
They answer customers, tenants, and patients; book the visit; take the request properly the first time; and hand off to a person when they should. On web chat, text, voice, and email.
See what first-line agents doThe work behind it
Systems your business runs on
Multi-step workflows, several agents working together, answers drawn from your documents and data, and the testing, controls, and records that make all of it safe to rely on.
See the production workWhat it takes off your plate
First-line work, done properly
- Answer the first question
- Tenants, customers, and clients get accurate answers from your own policies, leases, price lists, and manuals at any hour, with the source cited.
- Coordinate the schedule
- Book, confirm, and reschedule visits across technicians, vendors, and customers, and chase the missing confirmation so nobody else has to.
- Simplify the specification
- Turn product sheets, bid packages, warranties, and code requirements into plain answers a customer or field tech can act on.
- Take the intake properly
- Capture maintenance requests, service calls, and job details completely the first time, with urgency set by rules instead of guesswork.
- Keep internal work moving
- Agents for your own team that summarize threads, draft follow-ups, prepare checklists, and route approvals to the right person.
- Water is leaking through my bathroom ceiling and it’s getting worse.
- Safety First · Stay out of water near outlets or appliances. If it’s safe, shut off the main water valve.
- AssistantThanks for letting us know. Is the water coming from around the light or the vent? A photo would help the team too.
- Around the vent. Photo attached.
- Raised from Urgent to Emergency by the automatic safety check · Northgate Plumbing sent the job
Already at work in BrassTacks. Maintenance requests are taken by an AI intake assistant, and urgency is set by rules: a gas smell, a burst pipe, or no heat is raised to Emergency by a fixed safety check, never by a model’s guess; the tenant is shown safety steps written by people, not generated. Names in the sample are invented.
Who it’s for
Built for the people who answer the phone
On web chat, text, voice, and email, connected to your calendar and CRM.
Property owners & managers
Tenant questions about rent, leases, and building rules; maintenance requests captured and triaged; vendor visits coordinated around the tenant’s availability; showings scheduled.
Trades & service vendors
HVAC, plumbing, electrical, roofing, landscaping, cleaning. Booking and rescheduling, job updates to customers, quote intake, and plain-language explanations of equipment, warranties, and options.
Medical & dental offices
Scheduling and reminders, intake paperwork, and insurance and billing questions, with anything clinical always routed to your staff.
Every other business
Client intake, support and FAQs, appointment coordination, and answers drawn from your own documents, for professional services, small businesses, and internal teams.
How we build them
Built to be trusted with your customers
- Your rules decide
- Anything involving money, eligibility, safety, or a commitment runs through rules you approve. The model drafts; your rules decide.
- It knows when to hand off
- Clear escalation to a named person, with the conversation attached. Some topics it never handles on its own.
- It answers from your sources
- Grounded in your documents, price lists, and policies, with the source cited. When it doesn’t know, it says so.
- It says it’s an assistant
- Always identified as an AI assistant. Never presented as a person.
- Every conversation is on the record
- Logged and reviewable, so you can see what it said, what it did, and how much time it saved.
- Costs you can see
- Usage metered and capped, so a busy month never becomes a surprise invoice.
How an engagement runs
We learn how your front line works first
The same discovery-first approach as all our work. Sometimes the right answer is not an agent at all, and we will say so.
- Step 1
Find the first line
We sit with the people answering the phone and the inbox, and list what arrives, how often, and what a good answer looks like.
- Step 2
Draw the boundaries
What the agent may answer, what it may do, and what always goes to a person, agreed in writing before anything is built.
- Step 3
Build and rehearse
Built against your real questions and documents, then tested on past conversations before it talks to a customer.
- Step 4
Launch, measure, hand over
Monitored from day one, tuned on what it gets wrong, and documented so your team owns it.
Beyond chat and voice
The work behind the conversation
The chat window is the visible part. Most of the value, and most of the risk, sits underneath it.
Agentic workflows
Multi-step work an agent carries from start to finish: match the invoices, flag what doesn’t reconcile, draft the follow-ups, and stop for a signature where a person should sign off. Every step is logged and can be replayed.
Orchestration
When one agent isn’t enough, several share the work under a coordinator that assigns tasks, keeps track of where things stand, and knows when to stop and call a person. We design the handoffs as carefully as the agents.
Answers from your own information
Retrieval over your contracts, manuals, tickets, and databases, with the passage it relied on shown alongside the answer. We tune what gets retrieved as closely as what gets said; most wrong answers start with the wrong document.
Production LLM systems
The parts a demo leaves out: rate limits, retries, cost caps, caching, a fallback when a provider goes down, and a clean way to change models without starting over. Built to run at three in the morning with nobody watching.
Evaluation
A test set built from your real cases and scored before every change ships, so “it seems better” becomes a number you can compare. When a model update moves the score, we see it before your customers do.
Guardrails
Checks that sit outside the model, on what goes in and what comes out: topics it must pass to a person, claims it may not make, actions it may not take without approval. Asking a model nicely to behave is not a control. These are.
Integration with your systems
Connected to where the work lives: ERP, CRM, accounting, ticketing, document stores, data warehouses, and the older systems nobody wants to touch. Each agent gets access to what it needs and nothing more.
Governance
Who approved what the system may do, which model and prompt answered a given question, what it cost, and what changed since last quarter. Written down, kept current, and ready when an auditor or your board asks.
The standard we hold
Built the way software is built when it can’t be wrong
Where software can hurt someone, nobody ships on the strength of a good demo. Requirements come before code. Every requirement has a test that proves it. Every way the system could fail is written down, along with what stops it, and every change leaves a record. We build AI the same way. It takes a little longer at the start and saves a great deal later, because you can change the system without wondering what else you broke.
- Requirements first
- What the system must do, and what it must never do, agreed in writing before anything is built.
- Risks named and answered
- Each way it could fail, how likely and how serious, and the specific control that stops it.
- Traceable answers
- Any output can be followed back to the data, the model version, and the rules that produced it.
- Verified, not assumed
- Each requirement has a test that proves it, run again whenever anything changes.
- Changes on the record
- Nothing reaches production without review, a passing evaluation, and a note of what moved and why.
Business and technical
One team from the business case to the code
AI projects tend to stall in the gap between the people who know the business and the people who know the technology. One side can’t say what the system should be held to; the other can’t tell which mistakes actually cost money. We work on both sides of that gap. We start with the result you’re paying for, write it down as something that can be measured, and build back from there.
For the business
The case for doing it, what it will cost to build and to run, the risks, and what success looks like as a number.
For your technical team
Architecture, data access, security review, and a straight answer on what today’s models can and can’t be trusted with.
For both
One set of requirements everyone has signed, and progress reported against it in terms both sides can read.
Decks and demos
AI-washed decks are easy to make
A lot of what’s sold as AI right now is a slide deck and a demo built on examples chosen to work. It looks good in the meeting. Then it meets your real data, the odd cases start to pile up, and the firm that sold it is already pitching someone else. We’d rather be judged on what’s still running six months later: what it handles, how often it’s right, what it costs, and what it saved you.
What we measure
Results you can check
We agree the measures with you before we build, then report against them after launch. Usually some version of these:
- How much of the work it handles end to end, without a person stepping in
- How often it’s right, scored against answers your own people have checked
- How long a request takes from arrival to done
- What each handled case costs, model and hosting included
- When it hands off, and whether those were the right calls to hand off
Questions
What people ask us
- Will it replace my staff?
- It is built to take the repetitive first line so your people can spend their time on the work that needs them. Where those hours go is your decision.
- What happens when it gets something wrong?
- It is designed to hand off rather than guess, every conversation is logged, and anything with money or safety attached is governed by rules rather than the model. Reviewing and correcting mistakes is part of the engagement.
- Which AI model do you use?
- We pick the model that fits the job and budget, and can change it later without rebuilding your agent.
- Which channels and systems does it work with?
- Web chat, text, voice, and email, connected to your calendars and CRM. We add channels and integrations as clients need them, so if yours isn’t listed, ask.
- Where does our data go?
- Your agent works only from the information you connect for it. Most of what our AI-assisted features do is fenced in by fixed, deterministic rules: what the model is shown, what it may say, and what it may do. Anything involving money, safety, or eligibility is decided by those rules, not by the model. Hosting and retention details are set out in your agreement.
- Do you only work in real estate?
- No. Real estate is where we are deepest, and it shapes how we build. Trades, medical and dental offices, and other small businesses have the same kind of problem, and agents translate especially well.
- We already have a pilot that stalled. Can you pick it up?
- Often, yes. We start by measuring what it actually does on your data, which is usually the step that got skipped. Sometimes the fix is retrieval or testing rather than a new model; sometimes it’s a smaller scope. Either way, we’ll tell you which.
- Is this only for large companies?
- No. The discipline scales down. A ten-person firm gets the same written requirements and the same test set, sized to the job and the budget.
- Can you work with our IT and security team?
- Yes, and we’d rather start there. We document data flows, access, and retention in a form a security review can work through, and we build to your policies rather than around them.
Tell us what you need AI to get right
Describe the work, the systems involved, and what a mistake would cost. We’ll tell you what we’d build, how we’d prove it works, and whether it’s worth doing at all.
Or email info@brassicaworks.com