Questions, answered.
Everything prospective customers ask us on a first call — accuracy, speed, tooling, data handling and how we price. If your question isn't here, ask it directly. We answer the awkward ones too.
What we are.
What is human-in-the-loop, and what does “as a service” mean?
Human-in-the-loop means a trained person sitting inside an automated process, doing the part the model can’t do on its own. Labeling the data it learns from, approving the outputs it produces, judging whether its answers are right, and capturing data that doesn’t exist yet. “As a service” means you don’t recruit, train or manage those people. You brief us, and the work comes back done.
What does Akmon Data actually do?
Four things. We label your training data. We put reviewers between your model and your customers. We measure how accurate your model really is. And we go out and capture data you can’t buy. Most customers start with one and add others.
Who is Akmon Data for?
Teams building an AI model or an AI product: vertical SaaS companies with a model inside the product, data companies who need a delivery layer underneath them, small research labs with no operations function, robotics and physical-AI teams, and biotech teams who need specialist judgment rather than generic labeling.
Isn’t this just outsourcing?
No. Staff augmentation sells you seats and leaves the outcome to you. We own the outcome — the accuracy target, the throughput, the QA — and give you one point of contact. The work is also different in kind: rubrics that change weekly, genuine judgment calls, model evaluation. That isn’t back-office processing.
Why do AI teams need this now?
The easy data is gone. Models have been trained on what’s on the public internet, and what’s left has to be made by people: expert judgment, real-world video, edge cases, and honest scoring of whether a model got something right. At the same time AI moved from demos into production, which makes checking it a permanent need rather than a one-off project, and it moved into the physical world, where there is nothing to scrape at all.
The four lines.
What kinds of data do you label?
Images, video, text, audio and 3D. Detection and segmentation, transcription and diarization, preference data with written rationales, golden sets and rubrics, and verification of existing labels — yours or your model’s.
What is live ops?
People working inside your live product. Reviewing what the model produced, handling the cases it can’t, and catching what it got wrong before a customer sees it. You can have us approve every output before it executes, or set a confidence threshold and route only the rest to us. We run 100,000+ outputs a day with capacity past a million, around the clock.
What is model audit, and when would I need it?
Our reviewers grade your model against your rubric, before launch and after every release, so you get a failure rate you can defend. You need it most when you’re using a model to grade other models — that only works if the grader has been checked against real people first.
What data can you collect that I can’t buy?
Data that doesn’t exist yet. Egocentric (head-mounted) video, gameplay recordings, wearable and sensor studies, document sets, and synced video, audio and sensor streams. Captured to your task spec, then either fed straight into the annotation pipeline or delivered raw.
Do you do teleoperation?
Yes. Trained operators driving robots remotely, handling the edge cases the robot can’t — and generating training data while they do it.
I work in robotics. Isn’t collection the only thing you’d do for me?
It usually isn’t. Physical-AI teams tend to have large backlogs of their own footage that needs labeling and QA. That is often the faster place to start, with collection alongside it.
Do you have datasets already built?
Yes, licensed non-exclusively — each one captured while running a real business. Among them: 380,000 chats from plastic surgery consultations, six years of our own sales activity as a reinforcement-learning environment, guest chat from 1,000+ short-term rental bookings, and 1.1M+ messages a month of Slack from a 300-person ops team. Custom datasets to your spec are built the same way.
How we hold 99%.
How accurate is the work?
99%+, guaranteed. If a batch comes in under 99%, we redo it at no cost to you.
How do you actually hold that standard?
Our own QA layer sits over every project, and every label passes multiple checks before it reaches your bucket. Work is delivered with the QA report attached, so you’re auditing our numbers rather than taking them on trust.
Can you prove it against a vendor we’ve already tried?
We can show you the pattern. One US computer vision company came to us because their previous vendor couldn’t hold accuracy at production standard. The first project ran at a 100% acceptance rate, 111,550 frames passed QA in a month against 100,000 required, and that pilot became four production programs with 100+ annotation specialists on it.
How do you handle work where “correct” isn’t settled yet?
That’s normal on AI work and we price and plan for it. The rubric is often still being written when the work starts. We expect to retrain mid-project, and we don’t treat a changed definition of done as a new contract.
Speed, scale and scope.
How fast can you really start?
Live within hours of kickoff, with the first production batch inside a day. Brief one ops lead in the morning and the team is working by the afternoon.
How many people can you put on it, and how quickly?
10–15 people straight away. An initial team is fully deployed in four to five days, and a 50+ person ramp takes about two weeks. We’ve stood up a 15-person team in eight hours and 30 reviewers over a weekend. What we can’t do is put 100 people on something tomorrow, and we’d rather say so now than miss it later.
What do you need from me to start?
Your schema or rubric, whatever documentation exists, and access to your tooling if you want us working inside it. If the rubric is half-written, send it half-written.
Is there a minimum size? We only need a small team.
No minimum. A 15-person project is a real project to us, which is not true everywhere — with the largest vendors, small engagements get the attention their invoice justifies.
What happens when our spec changes?
Nothing, commercially. No headcount lock-in, no utilization minimums, and rework is built into the contract. The spec will change, and we don’t charge you for changing it.
How are you different from the large data vendors?
Three things. We deploy faster and we absorb scope changes without renegotiating. We work with you directly rather than sitting hidden behind another vendor, so you know who is doing the work and can talk to them. And we’re small enough that your project matters to us.
How the work runs.
Do we work in your tools or ours?
Either. We work inside your platform if you have one and bring our own if you don’t.
How do you deliver the data?
In your format, to your bucket, with the QA report.
How do we communicate day to day?
Same-day, in the channel you already use. AI work moves at the pace of Slack, not weekly calls and ticket queues, and we staff for that.
Where is your delivery team based?
The Philippines.
Your data, handled properly.
Can you handle confidential or regulated data?
Yes. We have Data Processing Agreements with standard contractual clauses in place for EU and UK customers, and we’ve been through vendor due diligence with health-sector clients. If you have a specific control you need to meet, raise it early and we’ll answer it directly rather than at signature.
Will our data be used to train models?
No. Our terms include no-model-training-on-customer-data commitments.
How we price.
How do you price?
Five ways, and we pick the one that fits how you think about spend rather than how we do.
Cost-plus — pass-through of hourly staff cost plus a service fee. Suits customers who want transparency, and longer engagements.
Hourly — billed per hour worked. Suits QA, live ops, and anything where scope isn’t settled.
Credits — pre-buy a block of hours and draw it down. Suits project work, pilots and bursty volume.
FTE-based — a monthly rate per dedicated person. Suits steady work and customers who want a named team.
Output-based — per approved unit: per video minute, per approved hour, per observation. Used on collection and some annotation work, once throughput is confirmed.
What will it cost us?
That depends on volume, the quality bar and which model above fits. We’d rather understand what it takes to deliver than quote a number we have to walk back, so a first call covers the work and the quote follows it.
The work we turn down.
Is there work you turn down?
Yes, three kinds. Engineering staffing. Anything requiring someone on site in the US. And work as an anonymous subcontractor behind another vendor — if we’re doing it, you know we’re doing it.
Still have a question?
Tell us what you are working on. We start within hours of kickoff.
The last 1% is human.