Continua AI Review 2026: The Cited Investigation Workspace That Reads the Whole Data Room

Product details checked against askcontinua.com, the pricing, benchmarks, email agent, trust and API pages, and the terms and privacy policy on September 15, 2026. Screenshots are from the live site.

Quick answer: Continua AI is an agentic investigation workspace for professionals who have to reason over a big, messy pile of documents and defend the answer afterwards. You attach the pile, describe the investigation in plain language, optionally set what to look for and how to organize the report, and submit a long-running task. What comes back is a report where the material claims carry page-level citations you can click. The company's argument is that this is a different job from chat: instead of retrieving a few passages and answering from them, it claims to read the whole record, up to a 1-billion-token context, so the amendment that quietly kills a CIM claim is not skipped because a retriever decided it was unlikely to matter. Pricing is $49 a month solo, $99 a month for teams, enterprise on request, and the first tasks are free with no card. There is an inbox agent and a batch API. The team is small, the public footprint is thin, and the benchmark numbers are first-party, so this review spends as much time on what to verify as on what is promised.

Why "read everything" is the whole pitch

Anyone who has run a real diligence, a franchise disclosure review, a medical chronology or a federal proposal compliance pass knows the failure mode. The answer you need depends on a later amendment. The spreadsheet contradicts the deck. The event that matters is buried in year three of the records. The sentence that changes the deal only makes sense next to a clause forty pages earlier.

Most AI document tools in 2026 are built on retrieval-augmented generation (RAG). Your files get chunked, the chunks get embedded, and when you ask a question the system pulls the top-k most similar chunks and hands them to the model. That works for "what is the termination notice period" and fails for "where does the seller's story stop matching the documents." The second question has no single similar passage. It needs the whole record held in view at once.

Continua's positioning is built on that distinction. The homepage calls it an "Agentic Investigation Workspace for Professionals," the category line is "Cited Analysis for Long Document Sets," and the four "shared data layer" claims are explicit: multimodal inputs, "100% full-context reasoning" instead of top-k RAG, a "live fact universe" of events, relationships, conflicts and changes, and agent-native artifacts. The marketing numbers on the same page are a 1-billion-token context window, 100% of pages read end to end, 100+ citations per report, and zero documents retained after a normal task.

Those numbers are vendor claims. Nobody outside the company has audited them. But the thesis behind them is sound, and it is the right thesis for the work this tool is aimed at.

What Continua AI is, in one paragraph

Continua AI is a hosted workspace at app.askcontinua.com where you sign in with Google or an emailed magic code, attach a file pile or a persistent Project, write the investigation in plain language, and submit a long-running task. The system produces a cited report you can follow up on in a thread, export as PDF or DOCX, and share as a read-only link. Normal-task attachments are deleted when the task finishes. The same engine is reachable by email through [email protected] and by a REST API at app.askcontinua.com/api/v1. Target users are legal, financial, healthcare, regulatory, commercial and research professionals.

The workflow, step by step

The help pages and homepage describe one consistent flow:

1. Attach the pile. Individual files, a folder, a Project, or files picked from Google Drive (Docs, Sheets, Slides and ordinary files). The company says it only touches the Drive files you select.
2. Describe the investigation. A paragraph, not a command. The homepage demo prompt is a good model: "I've attached the data room for a business I'm under LOI to buy... Cross-check the CIM's claims against the actual financials and documents, and tell me where the numbers don't match, which contracts have change-of-control clauses, customer concentration risk, and anything the seller glossed over. Cite every finding to the exact document and page."
3. Set the controls, optionally. Finding Criteria (contradictions, flags, gaps, risks), Composition Criteria (how the report is organized), guardrails (Lite for broader discovery, Strict for conservative, evidence-backed output), and the model, where the plan allows.
4. Submit and wait. Tasks are asynchronous. You can ask for an email when the report is ready.
5. Work the report. Follow-up questions in a thread, visual dashboards (charts, KPI cards, relationship maps, timelines, tables), PDF or DOCX export, read-only share link. A Web Search toggle adds public web sources as extra citations.

The controls are the part that separates this from a chat window. "Strict" guardrails plus explicit Finding Criteria is how you tell the system that a missed clause is worse than a shorter report.

The prompt that gets you a usable report

The demo prompts on the site are better than most users' first attempt, and they share a structure worth copying. Give the system four things:

• The role and the clock. "Solo buyer under a 45-day exclusivity window." "Plaintiff's counsel preparing a demand package." "Proposal manager with a Section L page limit."
• The comparison. Investigation is almost always X against Y: CIM against financials, FDD against the agreement and addenda, draft proposal against solicitation plus amendments, treatment notes against the billing ledger.
• The threshold. "Every gap in care longer than 30 days." "Any customer above 10% of revenue." "Any change-of-control clause, including ones triggered by asset sale."
• The output shape. "One page." "Table of contradictions with document and page." "Date-ordered timeline, duplicates counted once."

Then a constraint sentence: "Cite each finding to the exact document and page. If a claim cannot be supported by the record, say so rather than guessing." That last line matters because the terms of service are blunt that reports and citations may be incomplete or incorrect. You want the guardrails on your side.

Two storage modes, and what "zero retention" actually means

This is the privacy story and it is worth stating precisely, because the phrase "zero document retention" is doing a lot of work on the site.

Mode	What happens to your files
Normal task	Attachments are temporary and deleted after the task completes. This is the default.
Project	You choose persistence so the same pile can be reused across questions and shared with a team. Team plan feature.

What is still kept in both modes: account data, task settings, history, the generated reports and their citations, billing, and diagnostics. "Zero retention" means the source files die with the task, not that the account is stateless. That is a reasonable design. It just is not the same as "nothing is stored," and a security questionnaire should say so.

The other commitments on the trust page are clear: customer documents, prompts, task data and settings are not used to train models by Continua or its AI providers; customer data is not sold; you own both the submitted material and the generated reports; account deletion on request, with some records kept for legal and accounting reasons and deletion within 30 days where applicable; payments through Stripe with no card data stored by Continua; no use under 18.

Pricing: unusually plain for this category

Legal and diligence AI is a category where "contact sales" is the default price. Continua publishes numbers.

Plan	Price	Who it is for	What you get
Free	$0, no card	Trying it on a real file set	First tasks free; changelog mentions two complimentary tasks per week
Professional	$49/month	Solo professionals running recurring investigations	Private workspace, full history, shareable reports, complete-document analysis, exact-page citations, PDF and DOCX export
Team	$99/month	Groups that investigate and decide together	Everything in Professional, shared rooms, approvals, team connectors, shared Projects, collaborative follow-up
Enterprise	Contact	Security, procurement, regulated data	Everything in Team, private cloud or VPC, SSO and audit controls, enterprise connectors, BAA, vendor review support

Things the pricing page says in smaller type: subscriptions cancel before the next renewal; some models require paid usage or credits; the September 2026 changelog describes a current offer of 2× monthly credits, premium models and priority processing on paid plans. The exact credit table is not published, so if you plan to run frontier models on 9,000-page matters every week, ask before you assume the flat fee covers it. Free customers get a 100 MB per-file cap as of the same changelog.

The email agent is the feature I would actually use

A lot of the people this product targets do not want another web app. Underwriters, credit analysts and litigation paralegals live in Outlook. Continua's answer is an inbox agent.

How it works, per the email agent page: create an account, then from that registered address send mail to [email protected]. The subject line becomes the task title. The body is plain language, no command syntax. Up to 100 attachments in PDF, Word, spreadsheet and similar formats. The cited answer comes back by email and the task also appears in the web workspace, so you can continue the thread there or share it. You can name an existing Project as a source or a destination, which needs a Team plan. Attachments follow the same zero-retention default unless you put them in a Project.

The use cases listed on that page read like a sales map of who they are chasing: finance underwriting, credit review, contract review, due diligence, medical chronology, executive summaries, batch multi-document extraction, and knowledge graphs across attachments.

One practical note. Because the agent replies only to registered addresses, forwarding a borrower package from your work account means that work account needs to be the one on the Continua login.

The API: you poll for a memo, you do not stream tokens

Continua also exposes a developer API (documented at askcontinua.com/api), and it is worth understanding what kind of API it is. This is a batch investigation API, not a chat completion endpoint.

• Base URL https://app.askcontinua.com/api/v1, bearer token auth, asynchronous, no streaming.
• Endpoints: POST /files to upload, POST /responses to start a task, GET /responses/{id} to poll, GET /responses to list.
• Statuses run queued, in_progress, then completed, failed or cancelled.
• Follow-ups chain through previous_response_id.
• An optional report object carries project_ids, findings_criteria, composition_criteria and guardrail_mode set to off, lite or strict, mirroring the UI controls.
• Up to 200 input_file items per response. Files are single-use unless they belong to a Project, and unused uploads expire.
• Output includes output_text plus structured findings with claim, quote, relevance and confidence fields, and billing detail.
• Idempotency keys are supported.
• Documented formats: .txt .md .json .pdf .docx .xlsx .csv .eml .msg .pptx .html .htm .xhtml. The homepage also markets audio and video.

The structured findings array is the interesting part for anyone building on top of this. A claim with a quote and a page reference is exactly what you want to pipe into a review queue where a human clicks through and approves or rejects each finding.

Model names in the examples are auto, deepseek-flash, gpt-astra and claude-opus, and the benchmarks page names Claude Opus 5, GPT 5.6 Sol and Gemini 3.7 Flash. Some of those look like internal aliases. Check the live model picker before you write "GPT 5.6 Sol" into a procurement document.

The benchmarks, and how to read them

Continua publishes a benchmarks page with five internal "matters," five runs per model pair, a median of 412 pages per matter and a longest matter of 9,600 pages. Graders are described as blind. The scoring rule is the right one: a claim counts only if it is correct and the citation resolves to the page containing the supporting text. Unsupported claims score zero even when they happen to be true.

Benchmark	Task	Metric	Best result	Claimed lift vs same model alone
DD-Bench	Data room vs CIM	Contradictions found	Continua + Claude Opus 5, 91.2% ± 3.1	+38.6 points (52.6% to 91.2%)
FDD-Bench	Franchise disclosure vs agreement and addenda	Item-level gaps caught	Continua + GPT 5.6 Sol, 88.7% ± 3.6	+44.6 points (44.1% to 88.7%)
Chrono-Bench	Medical records to chronology	Events correctly dated and cited	Continua + Claude Opus 5, 93.5% ± 2.4	+35.3 points (58.2% to 93.5%)
RFP-Bench	Solicitation plus amendments vs draft proposal	Requirement coverage	Continua + GPT 5.6 Sol, 94.6% ± 2.1	+33.3 points (61.3% to 94.6%)
Cite-Bench	Amendment chains	Unsupported claims per report, lower is better	Continua + Claude Opus 5, 0.4 ± 0.2	94% reduction (6.8 to 0.4)

The method section is honest about what the comparison is. "Vanilla" runs get the same files and prompt, chunked to fit the model's context with standard retrieval. Continua runs get the identical pile through its full-context pipeline. Same weights, same prompt, same grader. The only variable is what the model can see.

What that means for you:

• The direction is credible. Full-record reasoning with citation-resolves-or-zero grading should beat top-k retrieval on exactly these tasks. That is the product thesis and the benchmark design matches it.
• The magnitudes are unverified. Five matters is a small sample, the 95% bootstrap intervals are noisy at that size, the file piles and rubric are not public, and no one outside the company has reproduced them.
• The baseline is not a competitor. The comparison is against Continua's own chunked retrieval setup, not against Harvey, Luminance, Spellbook, NotebookLM, or ChatGPT Deep Research with the same uploads. A "Continua vs the thing you use today" number does not exist yet.

So: a well-designed internal benchmark that tells you what the company optimizes for, not a third-party leaderboard.

Trust and compliance: what is stated, what is not

For a tool pitched at lawyers, lenders and health systems, the trust page is where the real evaluation happens. Here is the split.

Stated in first-party materials

• No training on customer documents, prompts, task data or settings, by Continua or its AI providers.
• Customer data not sold.
• Customer owns submitted material and generated reports; Continua takes only a limited license to operate the service.
• Temporary attachments deleted after a normal task.
• Account deletion on request, within 30 days where applicable, with legal and accounting carve-outs.
• HIPAA-ready workflows and a Business Associate Agreement on request. Covered entities are told to execute the BAA before sending PHI.
• Payments via Stripe.
• Subprocessor categories: cloud, AI processing, auth, email, payments. Vendors not named.

Not published as of this review

• Cloud region or vendor.
• SOC 2 or ISO 27001 report.
• Encryption-at-rest specifics.
• A Data Processing Agreement beyond the BAA language.
• Penetration test summary or bug bounty.

The honest framing: "designed to support HIPAA workflows, BAA available" is a contract and a design intent. It is not a completed audit. If your compliance team needs a SOC 2 Type II to onboard a vendor, ask for it directly, and expect that a 2-to-10 person company may not have one yet.

The terms are also refreshingly direct that AI reports and citations may be incomplete or incorrect, and that Continua does not provide legal, medical or financial advice. A tool that says that on its own terms page is being straighter with you than one that does not.

How it compares

Continua does not publish a comparison matrix, so here is the field as it stands in September 2026.

Category	Examples	Where Continua differs, as marketed
Legal and diligence AI	Harvey, Spellbook, Luminance, Kira, EvenUp	Full-set reading rather than retrieval, page-level citation as a scoring rule, published $49/$99 pricing, email agent
Long-context assistants	ChatGPT, Claude, Gemini with uploads; NotebookLM	Investigation controls (Finding and Composition Criteria, Strict guardrails), asynchronous long-running tasks, structured findings with quotes, batch API
Medical chronology tools	Record summarizers, PI demand-package tools	Chronology as one of several workflows, zero-retention default, BAA
GovCon proposal compliance	Dedicated RFP compliance AI	Amendment-chain tracing across Section L and M, requirement coverage scoring

The generic assistants are the real competitor for a solo professional, because they are already paid for. The fair question is not "does Continua chat better." It is whether, on a real 400 to 1,000 page pile, it catches the amendment the assistant misses and whether every citation lands on the sentence it claims.

Things to know before you rely on it

• Young public product. The changelog is a single month block dated September 2026, the copyright is 2026, and there is no dated history before that.
• Thin footprint. No named founder or executive. No funding disclosed. No press coverage, G2 or Capterra reviews, Product Hunt launch or Reddit threads about this product that we could find. The only offline trace is a LinkedIn post about a lightning talk at a Querit AI Search Afterhours meetup in Mountain View. For a B2B tool aimed at regulated buyers, that is a data point, not a disqualifier, but it means you are an early customer.
• Small team. LinkedIn lists 2 to 10 employees. Ask about support SLAs before putting a deadline-driven matter on it.
• Credits are opaque. Premium models cost credits and the table is not public.
• Model names may be aliases. Verify what "GPT 5.6 Sol" resolves to in the picker.
• Mobile. The changelog mentions a mobile portal redesign, but there is no native app for this product.

The hands-on test worth running

Free tasks with no card make this cheap to check. Here is the pass that would settle it for a diligence or legal buyer:

1. Assemble a real pack: a CIM or deck, a P&L spreadsheet, one lease, and one amendment that changes a material term.
2. Ask for contradictions between the deck and the records, plus every change-of-control clause, with document and page for each.
3. Click every material citation. Count the ones that land on the supporting sentence and the ones that do not.
4. Run the identical prompt with the identical uploads in ChatGPT, Claude and Gemini.
5. Score four things: time to report, citation hit rate, whether the amendment was caught, and export quality in DOCX.
6. Send the same pack to the email agent from your registered address and compare the response.
7. After a normal task finishes, try to reopen the source files. They should be gone.

If Continua catches the amendment and the citations resolve, the $49 is easy. If it does not, you have lost an afternoon and learned something about the category.

Verdict

Continua AI is a coherent, narrowly aimed product. It picks the right enemy, which is retrieval that silently drops the page that matters, and builds the whole system around a scoring rule that punishes a wrong citation as hard as a wrong claim. The investigation controls, the structured findings in the API, the inbox agent and the opt-in persistence model all point at real professional workflow rather than a demo. Pricing is plain, the privacy posture is stated clearly, and the terms do not oversell.

What it lacks is outside evidence. The benchmarks are first-party, the company is small and quiet, and the compliance story is a promise rather than a certificate. None of that is unusual for a product this new. It just means the burden of proof sits with your own test pack.

Best for solo and small-team professionals in legal, finance, healthcare and government contracting who regularly face a 400-plus page pile and need to defend the answer. Less compelling if your work is short documents, if you need a completed SOC 2 today, or if you already have an enterprise legal AI contract that covers the same ground.

The cheapest way to find out is to run a free investigation on Continua AI with a pile you already know the answer to. No card, and the report will tell you within an hour whether the citations hold.

FAQ

What is Continua AI (askcontinua.com)?

Continua AI is an agentic investigation workspace for professionals who work with large, mixed document sets. You attach contracts, filings, spreadsheets, transcripts, audio or video, describe the investigation in plain language, and get back a long-form report where material claims are cited to the exact document and page.

How much does Continua AI cost?

As of September 15, 2026, the first tasks are free with no credit card, and the changelog mentions two complimentary tasks per week. Professional is $49 a month for a private workspace, full history, complete-document analysis, exact-page citations and PDF or DOCX export. Team is $99 a month and adds shared rooms, approvals, team connectors and shared Projects. Enterprise is priced on request and adds private cloud or VPC deployment, SSO, audit controls and a Business Associate Agreement. Some models cost extra credits.

Does Continua AI keep my documents?

Files attached to a normal task are temporary and deleted after the task completes, which the company calls zero document retention. Files only persist if you add them to a Project, which is an opt-in Team feature. Continua still keeps account data, task settings, generated reports, citations and billing records. It states that customer documents and prompts are not used to train models by Continua or its AI providers, and that customer data is not sold.

Is Continua AI HIPAA compliant?

Continua says its workflows are designed to support HIPAA-covered work and that a Business Associate Agreement is available on request. It asks covered entities to sign the BAA before sending protected health information. That is a contractual and design commitment, not a public SOC 2 or HITRUST certification, so verify it in your own vendor review.

How is Continua AI different from ChatGPT, Claude or Gemini with file uploads?

The pitch is completeness. Most chat tools retrieve a small slice of your files with top-k retrieval and answer from that slice. Continua claims to read the entire submitted record, up to a 1-billion-token investigation context, keep conflicts visible, and score a claim as correct only when its citation resolves to the supporting page. It also adds investigation controls like Finding Criteria, Composition Criteria and Lite or Strict guardrails, plus an email agent and a batch API.

What is the Continua AI Email Agent?

It is an inbox-native way to run the same investigation. From a registered address you email [email protected] with a plain-language request and up to 100 attachments. The subject line becomes the task title, the cited answer comes back by email, and the task also appears in the web workspace. Storing the attachments in a Project for follow-ups requires a Team plan.

Are Continua AI's benchmark numbers independently verified?

No. The benchmarks page reports five internal matters with blind graders and shows large lifts over the same model run with standard chunked retrieval, for example 52.6% to 91.2% on contradictions found in a due diligence data room. The file piles and rubric are not public, the sample size is small, and the baseline is Continua's own retrieval setup rather than a named competitor. Treat the direction as plausible and the magnitudes as vendor claims.

Summary

• Continua AI at askcontinua.com is an agentic investigation workspace: attach a large file pile, describe the investigation, get a cited report where claims resolve to a page.
• The differentiator is full-record reasoning instead of top-k retrieval, backed by investigation controls (Finding Criteria, Composition Criteria, Lite or Strict guardrails), an inbox agent at [email protected], and an asynchronous batch API with structured findings.
• Pricing is public: free first tasks, $49 Professional, $99 Team with shared Projects, Enterprise with VPC, SSO and a BAA. Premium models cost credits and the credit table is not published.
• Privacy posture is coherent: no training, no sale, normal-task files deleted after the task, Projects are opt-in persistence, BAA on request. SOC 2, cloud region and a DPA are not published.
• Benchmarks show 88 to 95% scores and 33 to 45 point lifts over the same model with chunked retrieval, but they are first-party on five matters. Run your own pack before you rely on it.

If you have built a tool in this space, SaaSCity lists SaaS startups on a free, human-reviewed directory with a live city map, and submitting a listing takes a few minutes.