Basis vs Synthetic (2026)

Agents code bank and credit card activity, automatically push high confidence transactions and flag low confidence items for review, and compare each client's ledger with statements and schedules. They reach the customer's systems through revocable tokens instead of stored credentials. The demo request form asks for the firm's current ERP, but Basis names none of the ERPs, ledgers or file systems it connects to.

SourceBasis, getbasis.ai/solutions/cas-and-advisory and getbasis.ai/securityread 2026-09-06

Customers connect bank accounts, a payroll system, business tools and email, each named only by type. The one named route is the Google Workspace and Gmail integration, which reads message content through Google's APIs for transaction categorization and reconciliation under Google's Limited Use policy. Connected Financial Information also covers bank and card feeds and accounting system data. Google Workspace is the only integration Synthetic names, there is no connector catalog, and the books leave by export, not by writing back to another system.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

The agents run the month end close end to end for every client using the firm's policies, templates, mappings and review standards. That close covers coding transactions, reconciling each client's ledger against statements and schedules, surfacing exceptions and drafting reporting packages and variance analysis.

An end to end tax platform deploys agents across the entire return, from intake to review ready workpapers, and for audit the agents carry out testing with cited evidence. Agents carry out multi step work across whole workflows on Basis's own platform and coordinate tasks across accounting processes.

SourceBasis, getbasis.ai/solutions/cas-and-advisory and getbasis.airead 2026-09-06

One fixed loop carries the work, and Synthetic's agent runs it. The customer connects everything, and the agent reads the business context, reconciles, asks clarifying questions and produces accrual basis books. The agentic systems also initiate workflows and provide workflow and task management. The customer composes nothing. There is no builder, no runtime and no configurable sequence.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

The agents operate continuously in the background across a firm's clients, coordinating tasks and returning completed deliverables for review instead of running only when invoked. Recurring closes run month after month, with high confidence transactions pushed automatically. Basis does not say what starts each run, whether a schedule or a ledger or bank feed event.

SourceBasis, businesswire.com release of 24 February 2026 (vendor issued) and getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

The agentic systems start work on their own. They initiate workflows and contact the customer where appropriate to request clarification or more documentation, without routine human review. The customer connects accounts once and the agent comes back when it is done, as an ongoing service working over data as it arrives. Synthetic names no schedule the customer sets, no event types from connected systems, no webhook and no channel beyond the application and its outreach.

SourceSynthetic, synthetic.ai/privacy-policy and synthetic.airead 2026-09-07

Agents apply the firm's client specific policies, templates, mappings, guidelines and review standards to recurring work. Approved procedures and prior work come into each recurring workflow month after month and year after year, so the firm's methodology and each client's context persist across runs and stay in use. Basis does not say how that knowledge is stored.

SourceBasis, getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

The agent reads the context of the business from connected banks, payroll, business tools and email. Connected data also includes the chart of accounts, corporate ownership structure and related documentation, and Gmail content is read for categorization. The accrual books it maintains are cumulative, with each period building on the last. No query surface, corpus description or retrieval mechanism is published.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

Each recurring workflow picks up approved procedures, prior work and accountant feedback month after month and year after year, which Basis calls carrying context forward. Basis describes no conversational session state, and it does not say where the carried forward context is kept or for how long.

SourceBasis, getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

The ongoing reconciled ledger is the application's own accounting data. Gmail message content is generally processed in memory without storage beyond the feature, and is kept only case by case. One learning clause covers improving models that serve only the specific user whose data is processed. Synthetic names no memory layer for its agent.

SourceSynthetic, synthetic.ai/privacy-policyread 2026-09-07

Agents push high confidence transactions automatically and route low confidence items to human review. Reviewers work in one inspectable platform that holds source support, calculations, explanations, confidence levels and exceptions. The firm's team resolves exceptions and applies judgment before deliverables go out. Basis does not say whether the confidence threshold can be configured.

SourceBasis, getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

Clarifying questions are the one point where a person comes in. The agent comes back with sharp clarifying questions when the business context is ambiguous and consults the customer before it proceeds. Otherwise the agentic systems categorize, post and reconcile autonomously, without routine human review by Synthetic.

Synthetic personnel do not independently audit, review or verify the correctness of outputs, and the customer decides at their own risk whether to rely on them. There is no approval step before posting, so the customer only decides afterwards. Where law requires it, customers have a legal right to human review, which is not a feature of the product.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

Basis holds SOC 2 Type II, ISO 27001 and ISO 42001 certification, and the SOC 2 report is available under NDA. Customers are kept strictly separate from one another, with role based access. Multi factor authentication is enforced across all accounts, and access to customer systems is held as revocable tokens instead of stored credentials. Basis names no SSO, SAML or SCIM through the customer's own identity provider and describes no admin permissions. GDPR and CCPA are marked coming soon.

SourceBasis, getbasis.ai/securityread 2026-09-06

Synthetic names no certification or attestation of any kind, and no SSO, roles or permission model. It commits only to commercially reasonable safeguards and expressly says security cannot be guaranteed. The company is at waitlist stage, with a registered office at an apartment address. No security program for general availability is published.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

Every agent action is logged and auditable, and AI steps show their sources, changes and rationale. Agents keep source support, calculations, explanations, confidence levels and exceptions in one inspectable platform so reviewers stay in control. Basis does not say whether the log can be exported outside the platform.

SourceBasis, getbasis.ai/security and getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

Customers get full access to their financials and the insights behind them. The service prepares management reports and dashboards and generates derived artifacts, including categorizations, reconciliations, anomaly flags and suggested entries. Synthetic names no log of agent actions, no trace of what was posted and why, and no run record a customer can open. Its personnel do not review outputs, so there is no QA record on Synthetic's side either.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

Hosting is US only, and all customer data is stored in the United States. Client data stays within US based infrastructure. Customers cannot choose a region, and Basis names no customer cloud, on premises or customer tenant option. Backups are encrypted, taken daily and kept with geographic redundancy within Basis's own infrastructure.

SourceBasis, getbasis.ai/security and getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

Visitors and customers connect through and to servers located in the United States, and information is kept on web servers and systems within the United States. International transfers are treated as consent to US processing. Customers cannot choose a region, and there is no region list and no customer environment option.

SourceSynthetic, synthetic.ai/privacy-policyread 2026-09-07

One platform is sold through practice area solutions for CAS and advisory, tax, audit and corporate accounting, and each solution has agents doing that practice's work. Basis calls it one platform for all of CAS. The agents are unnamed, so there is no named set of prebuilt agents for a customer to choose among. An end to end tax platform deploys agents across the entire return.

SourceBasis, getbasis.ai and getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

The product is one bookkeeping agent built for a single purpose and one vertical, software startups. There is no library, catalog, template or second agent to adopt.

SourceSynthetic, synthetic.airead 2026-09-07

Customer data never trains any model. Basis does not say which model providers sit behind the agents, and it names no way for customers or admins to select a model. Its blog carries a post on progress in OpenAI's models.

SourceBasis, getbasis.ai/securityread 2026-09-06

Behind the agent is Synthetic's own stack, which it calls a new kind of AI with tools built specifically for the job. Unnamed third party cloud infrastructure supports its AI models, and training on Gmail data is limited to models that serve that one user. No model provider is named, and Synthetic names no choice of model for customers or admins.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

The product is sold by demo. Basis publishes no docs, developer or API page, and it names no API, SDK, MCP server or agent card that an outside caller could use to drive it.

SourceBasis, getbasis.ai and getbasis.ai/securityread 2026-09-06

No API, SDK, MCP surface or developer documentation is published. Synthetic frames an embeddable, Stripe like service as its aspiration and long term vision, so it is a plan, not a product. Books are exported as a download, not through a callable interface.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

Confidence levels on each coded transaction send low confidence items to review, one output at a time. The company's security controls go through an annual audit. Basis names no evaluation harness, scored test cases or optimization loop, and no quality check for changes to an agent.

SourceBasis, getbasis.ai/solutions/cas-and-advisory and getbasis.ai/securityread 2026-09-06

No harness, test set, readable result, release gate or live metric on the agent is published for customers. Synthetic works with design partners to validate reliability, and its own development practice includes quality assurance and testing. The founder's public standard is that nothing releases until it is more reliable than a human bookkeeper, and Synthetic applies that standard itself before release.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07

Agents reach the customer's ledger and files through revocable token access and work over documents inside Basis's own platform. Basis names no browser, desktop or computer control.

SourceBasis, getbasis.ai/security and getbasis.ai/solutions/cas-and-advisoryread 2026-09-06

Account connections carry the work, and the agent reads email content through the Google Workspace APIs. It does not drive a browser or operate a computer. Synthetic names no interface control and no open research on a case at run time.

SourceSynthetic, synthetic.ai and synthetic.ai/privacy-policyread 2026-09-07