Case study: the agents running Rye Valley every week, and the $425K they save

AI agents work in production at a small company, and we can show you a live install instead of a demo. Rye Valley, the holding company behind Hockey Stick, runs agents every week on top of a company brain, and within about a month of the install, it cut $425,000 in annual expenses and the team now ships 17x faster. This page walks through what runs and what we learned keeping it all alive.

Why we publish our own install

Most writing about AI agents is a tool roundup: ten products somebody tested for a week, scored and listed. That tells you which products exist. It doesn't tell you whether any of this survives contact with a real company's calendar, inbox and CRM for months at a time.

So our case study is the install we run ourselves. Rye Valley operates on a company brain with agents on top, built the same way we build for clients, and it has been carrying real work week after week since launch. We're careful about what that proves. One install is one install, and our team built the system it runs on, which is an advantage a client doesn't start with. What it does prove is that the pattern holds in production: agents on top of a current memory, scoped to repeat work, checked by humans, running on a normal company's normal tools.

You can see what the brain records for yourself. Michael Houck runs Rye Valley, and his public decision log at michaelhouck.co/decisions is kept in it: each hard call, dated, with the reasoning behind it.

The agents that run every week

The install runs the same first set we build for clients, each one wired into the tools the team already uses.

The morning briefing. Each weekday morning, the person running the company gets one briefing: what changed overnight, what's due and what needs a decision.

Sales follow-ups. After sales calls, a digest of the follow-ups owed and the next steps lands, built from the call transcripts and the history with each company, so nothing depends on someone's memory.

Call prep. Before a call, a short prep sheet arrives: who's in the room, the history, what was promised last time, the open items. Nobody spends the first five minutes reconstructing the last call.

The weekly scoreboard. The weekly numbers assemble themselves from the systems they live in. The team reads the scoreboard instead of building it.

Invoice chasing. Unpaid invoices get polite, persistent reminders on a schedule. The agent never forgets a follow-up and never feels awkward sending the third one.

Support reply drafts. Replies to support tickets are drafted from past support history and queued for a person to approve. A person stays on every send.

The answer layer. Anyone on the team can ask the brain a question from Slack or a text message and get the answer in seconds, with the meeting or document it came from. This is the agent that absorbs the repeat questions that used to interrupt whoever held the answer in their head.

Every one of these shares two properties. The work repeats, and the output can be checked against a source in under a minute. Those two properties are our test for whether a workflow belongs with an agent at all, and we'll tell a client when a workflow they want automated fails it.

What that work is worth in staff time

We price staff time one way everywhere, the same method our audit tool uses, so our numbers always agree: hours a week on the workflow x the hourly rate of the person doing it x 52 weeks. The rates come from the audit engine: a founder's hour at $150, sales at $75, finance at $70, operations at $55, support at $45.

Here's how that method prices the workflows above in a worked example. The weekly hours below are illustrative placeholders, there to show the shape of the math rather than to report our own timesheets. Your hours will differ, and so did ours.

Workflow an agent carries	The math (illustrative hours)	A year of staff time
Sales follow-ups drafted	6 hours a week x $75 x 52 weeks	$23,400
Meeting briefs prepared	2 hours a week x $150 x 52 weeks	$15,600
Weekly scoreboard built	4 hours a week x $55 x 52 weeks	$11,440
Invoices chased	3 hours a week x $70 x 52 weeks	$10,920
Support replies drafted	8 hours a week x $45 x 52 weeks	$18,720
Repeat questions answered	5 hours a week x $150 x 52 weeks	$39,000

Stack those six example lines and the method prices the set at $119,080 a year in staff time. The method caps any single workflow at 20% of the team's total weekly hours, so no one line can swallow the estimate, and it counts staff time only. It ignores the deals that stop slipping when follow-ups go out the same day and the knowledge that stops leaving when people do.

What the $425,000 is

The $425,000 is the annual expenses Rye Valley cut in about a month after the install, with only a fraction of it going back into AI costs. The cuts were possible because, for the first time, everything the company did and paid for sat in one place where it could be seen.

Alongside the dollars, the team shipped 17x faster, and we measure that in code. Across every Rye Valley repository, the 75 days after the install saw 4,349 commits, against 251 in the 75 days before, not counting automated log commits. We credit two things the brain changed. Decisions stopped getting re-litigated, because the reasoning from the first time was written down and findable, so settled questions stayed settled when the cast of a meeting changed. And the repeat work moved to agents, which returned hours to the people doing the shipping.

The memory underneath, and why the agents need it

None of the agents above would survive a month without the company brain underneath them. The brain is a single searchable memory of every meeting, every customer conversation and every decision with its reasoning, kept current by feeds rather than by anyone remembering to file things. Answers come with their sources attached.

The agents draw on that memory for everything they produce. The follow-up digest is useful because it draws on every prior call with each company. The support agent drafts a credible reply because it can search past tickets. The briefing agent knows what was promised last time because the promise is in the record. An agent without that memory writes generic email and guesses at answers, which is why bolting an agent onto a company with no searchable record produces demos that impress and tools that go unused.

The memory also does a job the agents can't, which is keeping knowledge from walking out the door. In most 5 to 200 person companies, operating knowledge sits in a few heads, in Slack scrollback and in SOPs that were accurate the week they were written. A wiki is supposed to fix that, and it does for about a month, until updating it becomes one more job nobody owns. Rye Valley chose plain files over a wiki for that reason. The brain fills itself from feeds, so the record stays current without anyone changing how they work, and the SOPs live next to the conversations and decisions that explain them, where drift is visible and the fix is one edit away.

What broke, and what still needs a human

Running the install taught us things a demo never would have, and they're the lessons we now build into every client system.

Feeds drift, and they drift quietly. When a tool changes its API or an export shifts format, a feed can stop without an error anyone sees, and the agents keep producing output from a memory that's going stale. The failure mode looks fine on the surface, which makes it worse than a loud crash. The fix is monitoring the feeds themselves, so a stopped feed gets noticed before the output goes stale. We learned to treat the plumbing as the product, and every install now watches its own feeds.

Drafts stay drafts until a person approves them. Early on, the temptation with every agent is to let it send on its own once the output looks good for a few weeks. We kept humans on the send button for anything a customer sees, because an agent working from a thin source can draft with a confidence the source doesn't earn. Approval costs the approver a few seconds per send, where one wrong confident email to a customer costs far more than all of those seconds put together.

Agents fail politely at judgment. Ask an agent to decide whether to fire a customer, reprice a product or make an exception to policy, and it produces something that reads like a decision and reasons like autocomplete. We scope judgment out of every agent and route those questions to people, with the brain supplying the record they decide from.

Maintenance never ends. Models update, tools change, the company keeps producing new kinds of repeat work. An install that gets no attention still runs, and it runs a little worse over time as drift accumulates. This upkeep is what our optional retainer exists to cover, and it's also why we hand over documentation that lets a team run the system without us if they'd rather.

Where agents are the wrong answer

The pattern that works in production has edges, and we'd rather draw them here than on a sales call.

Agents are wrong for judgment calls, for anything that needs taste, and for any output with no source to check it against. That time stays human, and a vendor who counts it in your savings estimate is inflating the payback.

A build like this is also overkill for a team of two or three whose main need is drafting help. An off-the-shelf assistant seat covers that well, and we'd say so on a call rather than quote you a build. The build earns its price when a team is big enough that knowledge concentrates in a few heads, the same questions recur every week, and follow-ups slip because nobody has the time.

And an agent purchased without a memory underneath it is the most common way this money gets wasted, whatever the sticker price was. If a vendor prices the agent without pricing the memory and the feeds that keep it current, ask what the agent will be answering from.

What the same install costs

A build like the one running Rye Valley starts at $15,000, and the brain with its first agents goes live in the first month, usually sooner. Before you sign anything you get a fixed quote: one price and a written definition of done. We've published a full breakdown of what drives the price and what the market charges if you want the line items.

After the build, an optional retainer from $2,500 a month covers new agents and the upkeep described above. When the build is paid, the system belongs to you, in plain files on machines you control.

Every build carries a savings review. Sixty days after launch, we run a savings review from the brain's own records, and if it finds under $15,000 a year in savings you agree are real, you get up to 2 months of the retainer free.

When we ran the audit on 11 real companies, the 1 to 5 person teams landed between $53,000 and $116,000 a year in their top 3 workflows, the 16 to 50 person teams between $157,000 and $192,000, and the 51 to 200 person teams between $315,000 and $655,000, all of it a floor because it counts staff time only. If you'd rather see your own numbers than ours, the free Leverage Audit takes 2 minutes: you give a website, a team size and the areas you care about, and it estimates the 3 workflows where agents would save you the most, with the hours and dollars shown.

The agents carry the repeat work every week, and the memory keeps what the company knows from leaving. That's what runs at Rye Valley today.