Lookback Engine: AI re-review of historical transactions for AML remediation
A cheap first-pass AI review of years of past transactions and closed alerts. It produces the case files and audit evidence that banks under AML enforcement must hand to regulators.
Crowded
Written for: Small startup team
Next step
Call the managing director who runs BSA/AML lookback engagements at FTI Consulting or Guidehouse. Ask what they pay per reviewed case and whether they would pilot AI first-pass review on sanitised data from a past engagement.
Why now
What changed since Triage Desk: the Amex fine ($350mn, $13bn unreported) shows the costly exposure is suspicious activity buried in history, not only the live alert queue. Any firm with 'systemic' findings must run a lookback, and those are typically priced per reviewed case. The forged 'cashmere sewing fabric' invoice in the FT laundering story shows that the stated invoice description, read against the counterparty's actual business, is a signal a model can now check at scale.
What became possible
Claude Haiku 5.5 is a small model with tokens up to 90% cheaper. That makes it affordable to read millions of historical transactions and closed alerts with a model, which used to be a job for armies of consultants.
Two ways in
- Lookback first-pass reviewer
Ingests historical transactions and closed alerts. Clusters them by typology, drafts a disposition and narrative for each, and sends only ambiguous cases to human investigators. Every decision is logged for the regulator.
- Customer
- Forensic consultancies running lookbacks for banks and card issuers under consent order
- Model
- Per-case fee for each case reviewed, licensed to consultancies as a white-label tool
- Invoice plausibility check
Flags trade-based laundering by comparing what an invoice claims (goods, price, volume) against the counterparty's known business and trade norms.
- Customer
- EMIs and B2B payment fintechs that process cross-border invoices
- Model
- API priced per invoice screened, with a monthly minimum
Competition
The "SR 26-2" guidance that Openlayer cites (which it says extends SR 11-7 to LLMs and agents) could not be checked against a Fed or OCC source.
Full competition notes (14 competitors)
Of your three ideas, the Alert Triage Copilot is the one that's already taken, and that's what drives the "crowded" rating. Unit21 and Hummingbird, two of the case-management systems you planned to pull exports from, now ship native triage agents that recommend dispositions with visible reasoning. Hummingbird even offers its agents by API on top of other systems. Sardine, Taktile, Flagright and Hawk sell the same thing, and at least five startups were funded for it in 2025–26 (Greenlite, Variance, Tangos, Diligent, Vivox). An outside tool that ranks alerts from exported data starts behind every one of them. Being cheaper per alert won't make up for that, because buyers pay for being plugged into their existing workflow and for vendor-backed audit trails. There are three places where you could still win: 1. Vendor-neutral validation of other vendors' AI agents. This is the strongest opening. ValidMind and Openlayer are generic model-risk tools; I found no AML-specific test pack from either. Taktile publishes SR 11-7 guidance but only covers its own product. Sponsor banks have to approve the Unit21, Hummingbird or Taktile agents their fintech partners use, which is the oversight problem Unit21's BaaS partnership is aimed at. A product that takes a bank's labelled alerts, replays them through any vendor's agent, and produces an examiner-ready file looks open. That file would cover: agreement with analysts; ROC curves and threshold choice; Holm-corrected comparisons between model variants; drift monitoring; a challenger model built from your cheap decision model. The main risks are that ValidMind adds AML templates, or that the triage vendors bundle validation packs themselves. 2. On-prem distillation into a small model. So far this is only done as custom services work (Ertas, SumatoSoft, CLPS). A packaged way to move from the hosted API into a 2–3B model inside the customer's own cloud environment, with the validation file carried over, is still open. The buyers are banks with data-residency rules, but they buy slowly. 3. Positioning the cheap model as an independent challenger, not a copilot. Model-risk teams need an independent benchmark to test a vendor's agent against. Selling the decision model in that role avoids competing head-on with the triage vendors. My recommendation is to lead with the validation and challenger pack, sold to sponsor-bank and fintech model-risk teams. The triage copilot should then be a feature inside that product, not the wedge. Caveats: Most of the vendors' performance figures come from their own marketing. The "SR 26-2" guidance that Openlayer cites (which it says extends SR 11-7 to LLMs and agents) could not be checked against a Fed or OCC source.
- Unit21 AI Agent
Ships L1 alert triage agents that run fixed investigation tasks per alert (event analysis, adverse media, network, prior activity) and recommend a disposition with visible reasoning. Unit21 says this runs at 100+ institutions; Green Dot and Nexo are named, and Nexo is quoted at 57% automation. It has a no-code Agentic Task Builder. In March 2026 it announced a partnership for sponsor-bank oversight of BaaS programs. It also owns one of your planned data sources (https://www.businesswire.com/news/home/20260323835434/en/Unit21-Announces-Partnership-to-Power-AI-Driven-AML-Workflows-and-Sponsor-Bank-Oversight-for-BaaS-Programs).
- Hummingbird Research & Review Agents
Launched June 2026. The agents triage alerts, apply the institution's own policies, fill out cases for human approval and show their full reasoning. They are also offered by API on top of other case-management systems, so they compete with an 'export and score' layer even outside Hummingbird.
- Sardine (Agentic AI for AML)
L1 and L2 triage agents for KYC, sanctions, PEP and adverse-media alerts. They clear false positives and recommend a disposition with a rationale.
- Taktile Agentic AML
Agentic AML launched in February 2026, covering transaction monitoring and case management, with a claimed 75% cut in false positives. Taktile also publishes SR 11-7 model-risk guidance for AI in AML (https://taktile.com/articles/managing-ai-model-risk-in-aml). That makes it the closest match to the 'triage plus validation story' pitch.
- Flagright
AML platform that markets AI agents for 24/7 alert triage and investigation from L1 to L3.
- Hawk AI (AML Investigative Agent)
Agent that automates investigation steps on AML alerts. It is listed alongside Unit21 in comparison write-ups.
- Variance
Raised a $21.5M Series A in March 2026 for AI agents that automate fraud, AML and sanctions investigations.
- Tangos
Reported $20M seed in July 2026 for agents that investigate alerts raised by banks' detection systems. Sources disagree on details such as where it is based.
- Greenlite AI
Raised a $15M Series A in 2025 for KYC and AML agents sold to banks. Its pitch to banks leans on explainability and model-risk readiness.
- Diligent AI / Vivox AI
Seed-stage European startups building KYC and AML agents, both funded in March 2026. More new entrants in the same slot.
- ValidMind
Model-agnostic model-risk platform for SR 11-7 work. It auto-drafts validation documentation, runs comparison tests and ongoing monitoring, and has LLM test suites. I found no AML-specific test pack. This is the direct competitor to the Validation Pack Generator.
- Openlayer
Evaluation and monitoring platform that sells validation for LLMs and agents under SR 11-7, with runtime monitoring and drift.
- AWS + Snowflake Cortex reference architecture
Big-cloud blueprint for LLM-automated AML alert triage. Banks on AWS or Snowflake can build their own version from it.
- Ertas / SumatoSoft / CLPS (on-prem fine-tuned AML models)
Ertas sells fine-tuned local models for AML. SumatoSoft (https://sumatosoft.com/wp-content/uploads/2026/07/A-private-LLM-compliance-copilot-for-a-European-digital-bank-SumatoSoft-Case-Study.pdf) built a private Llama copilot for a European bank. CLPS fine-tuned small LLMs on-prem for a Chinese bank's AML review. Together these show on-prem small-model AML work is already sold as services.
Checked 2026-10-07
Risks
- Regulators expect a documented, validated method. AI first-pass review may be challenged unless every model decision is sampled and checked by humans.
- Consultancies own the bank relationships and may build this themselves.
- Data rarely leaves the bank, so you may need on-prem or VPC deployment and an open-weight fallback model.
- The new tokenizer reduces part of the advertised price cut.
Strong market pull and team fit; open space is the weak point (6).
tech 7 · pull 8 · space 6 · fit 8 · build 7
Exposed
Companies that could benefit or be threatened. News and context, not investment advice.
Momentum of its main topic
Sources
What happened next
- +7d · 14 Oct 2026 nothing yet. No launches, funding rounds, or significant coverage of AML lookback engines or invoice plausibility checking products were found with publication dates after October 7, 2026. The search window spans only seven days and turned up no relevant announcements in this short timeframe.
Status history
- 2026-10-07 Flagged as buildable