MagBot.ai builds AI agents for drug safety work that is still largely manual, starting with literature screening and medical-information enquiries. Each decision shows the exact source text it is based on, and results are delivered into your existing safety database.
Every marketing-authorisation holder must monitor the published literature for adverse reactions to its products. In most companies, this relies on saved searches, spreadsheets and reviewers reading abstracts, most of which contain no valid case. In September, one of the more technology-led PV service providers was advertising for a freelancer to screen German journals manually.
Once an article meets the four minimum criteria, a serious case must be reported within 15 days in the US, EU and India. A slow weekly review reduces the time available to meet that deadline.
EMA’s literature monitoring service covers 309 chemical and 100 herbal substance groups. All other substances, all local journals, and all US and Indian obligations remain the company’s responsibility.
Most AI tools store a verdict with a plausible explanation. When an inspector asks where the article supports that decision, there is no way to show whether it came from the article or from the model.
Every MagBot agent follows the same four-step process. A decision that cannot cite its source is never saved, and nothing reaches your safety database without human sign-off.
The agent runs the searches and monitors the sources defined in your SOP, such as Embase, Medline, local journals or your medical-information inbox. Each source document is stored and locked.
The agent assesses the minimum criteria and proposes a decision, quoting the exact text that supports it. The quote is verified against the stored document before the decision is saved.
Your reviewer works from a prioritised queue with the supporting evidence highlighted, and approves, edits or rejects each decision. Articles that require the full text are flagged, not discarded.
Human checkpointApproved cases are sent to your safety database with a complete record: source, quote, Day 0, instruction version, model and reviewer.
Fully traceableEach result is presented with the supporting text highlighted in the article. Switch the example to see how MagBot handles an abstract that does not contain enough information to decide.
The explanation may sound correct, but nothing links it to the article. Two years later, it cannot be verified.
The quote must match the stored article word for word, or the decision is not saved. Anyone can repeat the check.
A 54-year-old woman presented with jaundice six weeks after starting the anticoagulant for atrial fibrillation. ALT was 612 U/L (12 times the upper limit of normal), with no other hepatotoxic exposure found. The drug was stopped and liver enzymes returned to normal within four weeks.
When asked whether an article describes an identifiable patient, a model may answer “no”. Whether that means “there is no patient” or “the abstract does not say” determines whether a reportable case is ever reviewed again.
“No patient.”
The article is marked not relevant and is never reviewed again.
“The abstract does not say.”
The full text describes a 62-year-old man. A reportable case is missed.
In MagBot, a “no” must be supported by a quote from the text. If no supporting text exists, the article is marked undetermined, the full text is requested, and the article remains open until a reviewer closes it.
We started with literature screening because the source data is public and the work is highly repetitive. Further agents are being developed with design partners, in the order they need them. All agents share the same evidence layer, so each provides the same quotes, review step and audit trail.
Runs your search strategy weekly and classifies each result as a potential ICSR, relevant for aggregate or signal work, or not relevant. Each decision quotes the article.
Answers enquiries from healthcare professionals and patients using approved sources only, including the SmPC or USPI, patient leaflets and standard response letters, with the source passage shown.
Drafts the literature and evidence sections of PBRERs, PSURs and DSURs using screening results that have already been reviewed and cited.
Identifies potential adverse events in medical-information, call-centre and patient-support records, and requests any information a case still requires.
Reduces repetitive work for case processors, while your safety database remains the system of record.
Compiles the evidence your safety physicians and QA team need for their decisions, without making those decisions.
The work is similar across organisations: high volumes, fixed deadlines, and qualified staff spending hours on reading and data entry. What differs is who performs the work and how it is funded.
For mid-size and growing authorisation holders that screen in-house, often across the US, EU and India. MagBot performs the first review and your team makes the final decision, with a record that stands up to inspection.
Global and local sources classified as potential ICSR, aggregate-relevant or not relevant
Recorded for each article, so reporting timelines are visible from the start
Search strategy, results, decisions and sign-offs logged as work is completed
Reviewed findings carried directly into PBRER and PSUR literature sections
For CROs and PV service providers that screen literature for multiple clients. Per-unit pricing aligns with your own pricing model, and your reviewers retain control of every decision.
Separate search strategies, products and SOPs for each client
MagBot performs the first review, so your team can focus on assessment
Every decision can be exported with its quote for client audits
For medical-information teams handling enquiries from doctors, pharmacists and patients. Answers are drawn only from approved documents, each with its source passage.
Based on the SmPC or USPI, patient leaflets and standard response letters
Off-label or unsupported questions are passed to a team member with full context
In development: potential adverse events routed to your PV team
The medical rules are configurable. Wherever a decision must be justified from documents, the same evidence layer applies. We take on these projects only with a design partner.
Each decision linked to the policy and claim documents
Clauses flagged with the exact wording behind each finding
Where our conversational agent already handles live traffic
Most tools in this market describe their AI as proprietary, auditable or explainable. These are claims about the model. MagBot applies fixed rules around whichever model you use, so an auditor can verify each result independently, without relying on the model.
| Typical AI safety tools | MagBot | |
|---|---|---|
| How the AI is checked | A confidence score or a written explanation | An evidence gate. The quote must match the stored source, or the decision is not saved. |
| Negative findings | Saved as not relevant | Must cite supporting text. Otherwise, they are marked undetermined and remain open. |
| Performance claims | Self-reported efficiency percentages | A sealed evaluation set excluded from development, with recall measured on your own literature. |
| Traceability | A log of user actions | Full provenance. Each decision links to its source, quote, instruction version, model and reviewer. |
| Safety database | Their own, or bundled with one | Database-neutral. Works with the system you have already validated. |
| Where it runs | The vendor’s cloud and the vendor’s model | Your cloud tenancy or on-premises, with interchangeable commercial or self-hosted models. |
Links every decision to verbatim source text before it is recorded.
A third outcome alongside yes and no, so missing information in an abstract never results in a closed case.
A set of labelled articles excluded from development and used only to measure recall and precision.
Versioned prompts, rules and models, each tied to the decisions it produced.
Comparison based on competitors’ public materials, October 2026. Technical details are available under NDA.
All three regulators require literature monitoring. They differ in review frequency and reporting timelines, and your SOP must meet the strictest requirement that applies to you.
| Region | Regulator and instrument | Required review frequency | Reporting timeline |
|---|---|---|---|
| United States | FDA21 CFR 314.80 · 600.80 | All adverse experience information, including the literature, must be “promptly reviewed”. No interval is specified, so your written procedures must define it. | 15 calendar days for serious, unexpected cases |
| Europe | EMAGVP Module VI, Rev 2 | At least once a week in widely used reference databases such as Medline and Embase, plus local journals. | 15 days serious, 90 days non-serious |
| India | CDSCO and PvPIMAH Guidance v2.0 | Monthly review across PubMed, ScienceDirect and Scopus. This is guidance rather than a regulation. | 15 days serious, 30 days non-serious |
Sources: eCFR 21 CFR 314.80; EMA GVP Module VI Rev 2 and the EMA medical literature monitoring service; IPC/PvPI MAH Guidance v2.0. Verified October 2026.
Inspection-ready records. Search strategy, results, decisions and sign-offs are logged as work is completed, so no records need to be reconstructed before an inspection.
Supports your validation. Versioned releases, change control and the documentation your team needs for computerised system validation.
Most vendors focus on time saved. Recent regulatory guidance focuses on different questions: what performance was defined in advance, how it was tested, and where human control is maintained.
A risk-based framework for showing that an AI model is credible for its intended use.
Principles for AI in pharmacovigilance: predefined performance criteria, open test sets and human oversight proportionate to risk.
Most provisions of the Act now apply, including its transparency and human-oversight requirements.
99%
of valid cases appeared in the top third of the ranked list in Clarivate’s published study with GSK. This is the only public recall result we are aware of in this market, so it is our reference benchmark.
Not yet
Our sealed evaluation set is complete, but measurement is still in progress, so we do not yet publish a figure. Design partners receive recall measured on their own literature.
When our results are available, we will publish them together with the measurement method, whatever the outcome.
Safety data and unpublished case material cannot be sent to a vendor’s shared cloud. MagBot is deployed within infrastructure you already control and have assessed, and integrates with your validated safety database.
Agents gather evidence and prepare drafts. The following safeguards are built into the platform rather than left to configuration.
Pricing is based on articles screened or enquiries answered, with no platform licence fee. The model suits mid-size authorisation holders, and service providers can incorporate it into their own pricing.
Scales with your portfolio and search volume, with no large upfront licence.
Costs scale with enquiry volume, and human handoff is included.
Designed to improve margins on outsourced screening, while your team continues to perform the review.
MagBot.ai is built by steppingedge, a profitable product engineering company based in Coimbatore, India, that builds and operates production systems for enterprise customers.
Much of the effort in regulated AI lies in less visible areas: integrations, access control, review workflows and change control. Our team has delivered this work for many years.
Investors: we are not currently raising. MagBot is funded by steppingedge, and any future raise will be based on measured evidence.
A senior team that builds the agents directly, rather than managing delivery remotely.
Our conversational agent already handles live customer traffic every day.
Designed around the literature requirements in all three markets.
You work directly with the team building the product, with no account management layer.
We are looking for two or three marketing-authorisation holders or PV service providers. A pilot requires real data, your current workflow, and an experienced reviewer who will critically assess the output.
