For textbook & curriculum publishers
Every AI companion promises it stays inside your titles.This one publishes the score.
A study companion built from the titles you attach to it, running under your own name. It never types a quote itself, and we publish what that produced — the one failure included.
- 0
- quotes made up, across 1,323 checked against the documents
- 97%
- more of our benchmark's rubric checks passed than the same model with no books attached
- 1,000
- pages per source, up to 50 sources per subject, chosen by your editors
The risk on the record is that your edition reads like everyone else's.
Pearson lists its content and channel being “insufficiently differentiated in terms of outcomes or learner experience” among its principal risks. Pearson plc 2026 interim results (risk disclosure, 31 Jul 2026; not a measured loss) A disclosure is not a loss. It says where a board believes the exposure sits, in a document that has to be accurate.
- 01
Your editors choose what goes in
Not a crawler. Record by record, up to 50 sources per subject and 1,000 pages each, with creator, date and document type attached.
- 02
Scans become searchable
Pages with no text layer are transcribed one at a time, historical handwriting included. You approve the cost before any page runs.
- 03
The student asks in plain words
The companion searches the documents you attached and nothing else, then answers with the passage and a route back into the book.
“They begin at six in the morning, and I have seen them asleep on their feet at the frame.”
The AI never types a quote. It marks the passage; our system copies the words out of your document and throws away anything that is not a character-for-character match. Ask it something the title does not cover and it says the material does not address it.
The evidence
Your standards team will look for the failure rate first,so we printed it.
Answers limited to your books, citations back to the page, your branding on the screen: somebody has offered you all three. The difference is what happens when a reviewer goes and checks.
- 97%
- more of our benchmark's rubric checks passed than the same model with your books taken away
- 0
- quotes that differed from the document they were copied out of
- 1,323
- quotes checked, one at a time, against the page they came from
Humy internal benchmark, August 2026: one configuration, 1,250 questions. 58.2% of the rubric checks passed with your books attached, 29.6% without. One of those 1,323 citations opened nothing, and we report it rather than round it away.
What that does not cover: the sentences around a quote are still written by the AI. In the same benchmark, roughly one answer in eight carried an invented date or reference in that surrounding text. The quote itself is always real. That gap is precisely why the page it came from is one click away. Read the benchmark.
It ships as your product,in the channel you already bill.
Your name, your identity, no Humy branding in front of the student. Three ways in, and the one you pick has a commercial consequence.
In your digital edition
Beside the chapter it belongs to. Nothing for a student to download, nothing new for your platform team to host.
Off a code printed in the book
A phone camera opens the companion for that title. A printed programme gets a live surface without a reprint.
Inside your own platform
Drive it through our API, so there is no second destination for a teacher to remember.
Inclusive access, where the institution is billed rather than the student, grew 20% at Pearson over the half year and now carries 50% of its core US courseware business. Pearson plc 2026 interim results (six months to 30 Jun 2026; one publisher, US courseware only)
A companion inside your edition stays in that relationship. One a student signs up to separately does not, and it puts a second bill in front of somebody who has already paid you.
None of the five below is a reason to choose us. Every vendor your rights director has met answers them the same way. They are on the table so the conversation can move past them.
Nothing you supply trains a model
Not by us, not by the providers we run on. Student conversations are not used for training either.
No pooling between publishers
A companion searches the documents attached to it and nothing else. Your material never reaches another publisher's.
Withdraw a title and its records go with it
Once they are gone, the answers that leaned on them stop coming. There is nothing left to quote.
Where the data sits
The US or Sweden, on Microsoft Azure. Signed student data privacy agreements cover 16 US states, 15 through TEC plus California.
Aligned to the frameworks you sell against
AP, Common Core, IB, TEKS and eight further frameworks, or a grading guide your editorial team uploads.
One title, one course, one term.
- Week 0 · 30 min
Pick the title
One where the questions teachers ask are already documented in your support queue.
- Weeks 1–4
Build it
You supply text, PDFs and scans. Transcription is costed and approved before any page runs.
- Before a classroom
Test it yourself
Put your reviewers' questions to it and read every transcript. A thin chapter shows up as a refusal, not an invention.
- One term
One course, live
Real students, real coursework. Every conversation exports as PDF or CSV for your editors.
Where we have not been yet
We have not run a programme-wide deployment with a publishing house: of the 13 businesses paying for Humy today, none is a publisher. We will not describe a track record we do not have.
The nearest is Pearson's Tennessee Connections Academy, an online public school that chose Humy for its middle school students. A pilot will not settle whether this helps you win a state adoption either; that is a judgement on the book.
Bring one title to the call.
Bring your editorial lead and whoever owns the digital product. Thirty minutes is enough to scope a pilot, and we will say on the call if the material is too thin to answer from.
Priced against the title once we have seen it. We would sooner be a cost you absorb into an edition you already sell than a surcharge on top of one.
Questions from the editorial floor.
We build a source-grounded AI simulation of whatever a title is about: the person, the event, or the site at the centre of it. Each title becomes a set of source records your editors choose, and when a student asks a question the companion searches those records, answers from what it finds, and attaches the reference to the page it drew on. The student can open the passage and read it in place. The effect is that using the AI sends students into your content, not past it.
No. It can be a person, an event, or a place, and all three are built the same way: a name, a short description, and the documents your editors attach. A biography becomes the person it covers. A title on a treaty or a battle becomes that moment, which a student can question from inside. A title on a building or an excavation becomes the site, answering out of the records about it. Historical figures are the best-known case and the one our ready-made library is stocked with, but they are one case, not the limit of what your list can become.
Because a general tutor answers the curriculum, not your edition of it. Nothing in that exchange credits the sequence your editors argued over or the wording of a definition a teacher will mark against, and nothing routes the student back to a page. Pearson names its content and method of delivery being, or being perceived to be, insufficiently differentiated in terms of outcomes or learner experience as a principal risk in its 2026 interim results. A companion that can only answer with your material attached puts your edition back into the answer, and a published test gives your standards team something to check before they sign.
No. Your material is stored as source records that a companion searches at question time. It is not used to train models, not by us and not by the model providers we run on. Nothing from your documents is pooled with another publisher's, and you can withdraw a title and have its records removed.
Yes to both. Humy runs in the browser and can be embedded in your digital edition or learning platform by iframe, opened from a link or QR code printed in the physical book, or driven through our API. It runs as your product, under your name and your identity, with no Humy branding presented to the student. You stay the party of record with the institution, in the channel where the institution is already billed under your agreement.
Humy marks which passage it means, and our system copies the words straight out of the document, so the AI never types a quote itself. If the text does not match the document character for character, it is thrown away. In the benchmark we published, all 1,323 quotes were exact copies of the source, none highlighted the wrong passage, and none came from outside the documents attached to that figure. One reference out of those 1,323 pointed at no source at all, and we publish that too.
It can. The quote itself is copied out of your document and cannot be made up, but the text around it is still written by the AI. In the same benchmark, working from the documents cut made-up details from 22.6% of answers to 12.2%. That is about one answer in eight still containing an invented date or document reference. The cited passage is therefore always one click away, so a reviewer can check the claim against the page it came from.
Up to 50 source records per subject, and up to 1,000 pages per source: clean digital text, PDFs, scans with no text layer, images, and audio. Pages a computer cannot already read are transcribed one page at a time by Gemini 3.1 Pro, which also reads historical handwriting in facsimile material. You approve an estimated transcription cost before any page runs.
One title, one course, one term. We scope it with your editorial team, build the companion from that title's material, and you test it against the questions your reps hear from teachers before it reaches a classroom. Write to hello@humy.ai with the title you would start with and the audience it serves.
Yes, and it matters more there, not less. Because the companion answers from the adopted text, its answers stay inside the scope the adoption was approved against. Humy aligns assignments to AP, Common Core, IB, TEKS and eight further frameworks, and your editorial team sets the reading level and instructions per edition. It will not win you a state adoption: Texas reviews materials for alignment to the state standards, accessibility, and quality and suitability criteria, which is a judgement on the book itself.
Under the same terms as our school business. Student conversations are never used to train models, students can work through links or QR codes without accounts, and we hold signed student data privacy agreements covering 16 US states, aligned with FERPA and COPPA. Data is stored in the US or Sweden on Microsoft Azure. The signed agreements are published at humy.ai/security.
Something not covered here? Write to hello@humy.ai.