HerendilApplied AI, consulting and software
What could an AI agent do for your team? Work it out with your own numbers.
Herendil helps companies put agents to work on what they already know and do. We advise, and we build the software. This page is a small model of your case. Change any blue number and the rest recalculates. When it looks like your company, bring it to a call.
Drag a blue number sideways, or tap it for − and + buttons. Arrow keys work too.
Looking for answers
Your team has 40 people. Each spends about 5 hours a week looking for something the company already knows, in an old thread, a shared drive or one colleague’s memory. That adds up to 9,200 hours a year.
Assumes 46 working weeks a year.A company brain indexes what the company knows, decides and does, and lets people and agents ask it questions. Suppose it answers 40% of those searches well enough that nobody has to go looking. That gives back about 3,680 hours a year.
One repetitive workflow
Now pick one job that repeats, such as triaging support tickets or reconciling invoices. It runs 250 times a week and takes a person 15 minutes each time.
An agent does the first pass and a person checks it in 3 minutes. The job drops from 2,875 to 575 hours a year, a saving of 2,300.
Review can’t take longer than doing the job, so the model caps it there.- Looking for answers
- The workflow, or reviewing it
- Returned: 5,980 h
What inference costs
Each agent run reads and writes about 40,000 tokens. A question to the company brain uses about 20,000, and each person asks 6 a day. That is roughly 130 million tokens a month.
Most tokens are the documents an agent reads, which is why the counts look large.At $4 per million tokens, a blended input and output price, the bill comes to about $521 a month. One workflow run costs $0.16 in inference against $15.00 of a person’s time at $60 an hour.
Prices differ a lot by model and provider and fall often. Check the current price card.Where that leaves you
About 5,980 hours a year back, the work of 3.3 full-time people, for about $6,260 a year in inference.
Read every number here as an estimate. The model leaves out the cost of building, integrating and maintaining the system, which is usually larger than the inference bill. The 40% is a guess until someone measures it. Replacing guesses with measurements is the first thing we do together.
A full-time year here is 1,840 hours: 46 weeks of 40.How Herendil works
Find the hours
We sit with the people who do the work and measure where the time goes. Your model guesses 5 hours a week each. We replace that guess with a number.
Build one agent
We take the job that runs 250 times a week and build an agent that does the first pass inside your own tools. A person approves each result until the error rate earns more trust.
Connect the company brain
We index the places your knowledge lives and keep each source’s permissions, so people and agents only see what they are allowed to.
Measure, then decide
You see real hours, real error rates and the real bill next to the $521 a month estimate. Then you choose whether to extend it, hand it to your team, or stop.
Notes
- A company brain starts with permissionsWho may see a document decides what an agent may say about it.
- Estimating inference cost before you buildThe arithmetic behind section 3, with the inputs that move it most.
- Review time belongs in the estimateAn agent that saves twelve minutes and needs ten to check saves two.
- Where an agent should stop and askHanding work back to a person is part of the design.
Bring this model to a call
We will go through your numbers, mark which ones are guesses, and tell you whether an agent is worth building yet.
- Team
- 40 × 5 h/wk
- Brain answers
- 40%
- Workflow
- 250/wk, 15 → 3 min
- Tokens
- 130M/mo at $4/M
- Hours back a year
- 5,980
- Inference a month
- $521