Custom agents, built right.
When the library doesn’t have it, we build it. Then we operate it.
Per sprint: $32,000.
Our build approach
Scoping. Every custom agent starts with a written scope — what the agent does, what it reads, what it can write or send, who reviews its output, what success looks like in numbers. Scoping is one to two hours of structured conversation, not a multi-week discovery. By the end of the scoping call, you know what you’re getting and what we’re committing to.
Build sprints. Two-week sprints by default; four-week sprints for agents that integrate with three or more systems. The sprint includes design, build, integration, testing in a sandbox, user acceptance with a named pilot group, and rollout. We are not selling you a 90-day discovery followed by a 90-day implementation followed by a 90-day “operationalization.” We are selling you a working agent in your environment within 30 days of signing the sprint SOW.
Instrumented from day one. Every agent we build emits structured measurements. Inputs received, actions taken, outputs produced, errors, latency, override rate (how often the human reviewer changed the output), approval rate. The measurements feed your ABR. You will know how the agent is performing the day it goes live, not three months later when somebody finally builds a dashboard.
Human-in-the-loop by default. Agents draft. Agents propose. Humans approve. The default is review before send. We will discuss removing the review step for low-stakes workflows after we have three months of clean performance data, and only if you want to. We do not deploy agents that take significant actions unsupervised.
Owned and maintained by us. When the model changes, when the underlying systems change, when the data changes, the agent has to change with them. We do that work. Custom agents we build are not a one-time delivery; they are a continued operation. When something breaks at 9pm on a Tuesday, someone fixes it, and you have a phone number to call.
Example builds
A regional law firm: discovery response drafting
The problem. A 40-attorney plaintiff’s firm in Michigan was producing roughly 35 discovery responses per month, averaging four hours of senior associate time each. Response quality varied with the associate. The firm was passing on cases — losing matter intake to bandwidth, not capability.
The build. A custom agent in Microsoft Copilot Studio, reading from the firm’s iManage DMS via approved API, drafting first-draft responses against the firm’s discovery template library and the matter’s existing pleadings. The associate reviews, edits, and finalizes. The agent does not file; the partner still signs the response.
The result. Average drafting time dropped from 4 hours to 55 minutes. The firm’s monthly discovery response capacity grew without adding headcount. The quality variance narrowed because the template enforcement was tighter than memory across 12 associates. The pilot was 60 days; the workflow has been running for 8 months and is in the firm’s annual operating plan.
A 200-employee manufacturer: RFQ acceleration
The problem. A custom-machining shop in Central Michigan was losing about 15% of RFQs to slow turnaround. The estimator queue was 5 days deep in busy weeks. The CFO had no real-time view of RFQ volume vs. quote response time.
The build. A custom agent hosted in ACS Automate, monitoring the RFQ inbox and the customer portal, pulling prior quote history and current material costs from the shop’s ERP via a read-only database view. The agent drafts quote packages and routes them to the estimator in Teams. The estimator reviews the labor and margin calls, adjusts as needed, and approves.
The result. RFQ turnaround dropped from average 5 days to average 1.5 days. Quote win rate increased 7 percentage points in the first quarter post-deployment (multiple factors, but the team and the CFO attribute most of the lift to response time). The CFO now has a dashboard with RFQ volume, quote response time, and win rate by customer.
A regional CPA firm: engagement letter and SOW generation
The problem. A 12-partner CPA firm was generating roughly 80 new-client engagement letters per year. Managers were spending 2.5 hours per letter on what was largely template work. Inconsistency across managers produced compliance concerns at peer review.
The build. A Copilot agent inside the firm’s tenant, drafting engagement letters and SOWs against the firm’s master template library and the new-client intake form. The manager reviews and edits. The partner signs. The output goes to the firm’s document management system in the same folder structure they already use.
The result. Average drafting time dropped from 2.5 hours to 25 minutes. Letter consistency improved at peer review. Two managers reported the workflow let them take on a meaningful number of additional new-client engagements in the same number of weekly hours.
Where they run
Custom agents run in one of three places, picked at scoping based on what the workflow needs:
In your tenant. Microsoft Copilot Studio, Power Automate, or Azure OpenAI inside your Microsoft tenant. Best for workflows whose data stays inside Microsoft 365 and whose users live inside Teams / Office. Most of our pro-services builds run here.
In ACS Automate. Our hosted agent platform. Best for workflows that need to read across multiple systems (ERP + DMS + CRM, etc.) or that need a hosted execution environment outside your tenant. Most of our manufacturing builds run here, especially the ERP-integrated ones.
Hybrid. A user-facing surface inside your tenant (Teams bot, Copilot prompt) backed by execution in ACS Automate. Best for workflows that need both end-user familiarity and deeper-than-Microsoft integration.
We tell you which fits during scoping. If the answer is “it doesn’t matter, go with whichever you prefer,” we say that too.
Book an AI Readiness Assessment
Ready when you are.
Three weeks. Fixed fee. A real report and a 90-day roadmap. You keep it whether you hire us or not.