diff --git a/drivers/serial/Kconfig.stm32 b/drivers/serial/Kconfig.stm32
index c2a6f96c3b5..b8aaa9d764f 100644
--- a/drivers/serial/Kconfig.stm32
+++ b/drivers/serial/Kconfig.stm32
@@ -25,15 +25,34 @@ config UART_STM32
 
 if UART_STM32
 
-config UART_STM32U5_ERRATA_DMAT
-	bool
-	default y
-	depends on SOC_STM32U575XX || SOC_STM32U585XX || \
+if SOC_STM32U575XX || SOC_STM32U585XX || \
 		   SOC_STM32H562XX || SOC_STM32H563XX || SOC_STM32H573XX
+
+choice UART_STM32U5_ERRATA_DMAT
+	prompt "Workaround for DMAT errata on selected devices"
+	default UART_STM32U5_ERRATA_DMAT_LOWPOWER if PM
+	default UART_STM32U5_ERRATA_DMAT_NOCLEAR if !PM
 	help
 	  Handles erratum "USART does not generate DMA requests after
 	  setting/clearing DMAT bit".
 	  Seen in Errata Sheet 0499 § 2.19.2 and §2.20.1 for stm32u57x/u58x,
 	          Errata Sheet 0565 § 2.14.1 and §2.15.1 for stm32h56x/h57x
 
-endif
+config UART_STM32U5_ERRATA_DMAT_LOWPOWER
+	bool "Send first byte by polling"
+	help
+	  This option sends first byte via software polling and
+	  rest of the buffer is sent via DMA. This option is suitable
+	  for STOP low-power modes.
+
+config UART_STM32U5_ERRATA_DMAT_NOCLEAR
+	bool "Do not clear DMAT"
+	help
+	  This option keeps DMAT bit set. This may cause additional power
+	  consumption in STOP low-power modes.
+
+endchoice
+
+endif # U575 || U585 || H562 || H563 || H573
+
+endif # UART_STM32
diff --git a/drivers/serial/uart_stm32.c b/drivers/serial/uart_stm32.c
index 6728a7aa4dd..296085b8c8f 100644
--- a/drivers/serial/uart_stm32.c
+++ b/drivers/serial/uart_stm32.c
@@ -1434,7 +1434,7 @@ static inline void uart_stm32_dma_tx_enable(const struct device *dev)
 
 static inline void uart_stm32_dma_tx_disable(const struct device *dev)
 {
-#ifdef CONFIG_UART_STM32U5_ERRATA_DMAT
+#ifdef CONFIG_UART_STM32U5_ERRATA_DMAT_NOCLEAR
 	ARG_UNUSED(dev);
 
 	/*
@@ -1615,8 +1615,16 @@ static int uart_stm32_async_tx(const struct device *dev,
 	const struct uart_stm32_config *config = dev->config;
 	USART_TypeDef *usart = config->usart;
 	struct uart_stm32_data *data = dev->data;
+	__maybe_unused unsigned int key;
 	int ret;
 
+	/* Check size of singl character (1 or 2 bytes) */
+	const int char_size = (IS_ENABLED(CONFIG_UART_WIDE_DATA) &&
+			       (LL_USART_GetDataWidth(usart) == LL_USART_DATAWIDTH_9B) &&
+			       (LL_USART_GetParity(usart) == LL_USART_PARITY_NONE))
+				      ? 2
+				      : 1;
+
 	if (data->dma_tx.dma_dev == NULL) {
 		return -ENODEV;
 	}
@@ -1648,34 +1656,74 @@ static int uart_stm32_async_tx(const struct device *dev,
 	/* Enable TC interrupt so we can signal correct TX done */
 	LL_USART_EnableIT_TC(usart);
 
-	/* set source address */
-	data->dma_tx.blk_cfg.source_address = (uint32_t)data->dma_tx.buffer;
-	data->dma_tx.blk_cfg.block_size = data->dma_tx.buffer_length;
+	/**
+	 * Setup DMA descriptor for TX.
+	 * If DMAT low-power errata workaround is enabled,
+	 * we send the first character using polling and the rest
+	 * using DMA; as such, single-character transfers use only
+	 * polling and don't need to prepare a DMA descriptor.
+	 * In other configurations, the DMA is always used.
+	 */
+	if (!IS_ENABLED(CONFIG_UART_STM32U5_ERRATA_DMAT_LOWPOWER) ||
+	    data->dma_tx.buffer_length > char_size) {
+		if (IS_ENABLED(CONFIG_UART_STM32U5_ERRATA_DMAT_LOWPOWER)) {
+			/* set source address */
+			data->dma_tx.blk_cfg.source_address =
+				((uint32_t)data->dma_tx.buffer) + char_size;
+			data->dma_tx.blk_cfg.block_size = data->dma_tx.buffer_length - char_size;
+		} else {
+			/* set source address */
+			data->dma_tx.blk_cfg.source_address = ((uint32_t)data->dma_tx.buffer);
+			data->dma_tx.blk_cfg.block_size = data->dma_tx.buffer_length;
+		}
 
-	ret = dma_config(data->dma_tx.dma_dev, data->dma_tx.dma_channel,
-				&data->dma_tx.dma_cfg);
+		ret = dma_config(data->dma_tx.dma_dev, data->dma_tx.dma_channel,
+				 &data->dma_tx.dma_cfg);
 
-	if (ret != 0) {
-		LOG_ERR("dma tx config error!");
-		return -EINVAL;
+		if (ret != 0) {
+			LOG_ERR("dma tx config error!");
+			return -EINVAL;
+		}
+
+		if (dma_start(data->dma_tx.dma_dev, data->dma_tx.dma_channel)) {
+			LOG_ERR("UART err: TX DMA start failed!");
+			return -EFAULT;
+		}
+
+		/* Start TX timer */
+		async_timer_start(&data->dma_tx.timeout_work, data->dma_tx.timeout);
 	}
 
-	if (dma_start(data->dma_tx.dma_dev, data->dma_tx.dma_channel)) {
-		LOG_ERR("UART err: TX DMA start failed!");
-		return -EFAULT;
-	}
-
-	/* Start TX timer */
-	async_timer_start(&data->dma_tx.timeout_work, data->dma_tx.timeout);
-
 #ifdef CONFIG_PM
 
 	/* Do not allow system to suspend until transmission has completed */
 	uart_stm32_pm_policy_state_lock_get_unconditional();
 #endif
 
-	/* Enable TX DMA requests */
-	uart_stm32_dma_tx_enable(dev);
+	if (IS_ENABLED(CONFIG_UART_STM32U5_ERRATA_DMAT_LOWPOWER)) {
+		/**
+		 * Send first character using polling.
+		 * The DMA TX needs to be enabled before the UART transmits
+		 * the character and triggers transfer complete event.
+		 */
+		key = irq_lock();
+
+		if (char_size > 1) {
+			LL_USART_TransmitData9(usart, *(const uint16_t *)tx_data);
+		} else {
+			LL_USART_TransmitData8(usart, *tx_data);
+		}
+
+		if (data->dma_tx.buffer_length > char_size) {
+			/* Enable TX DMA requests */
+			uart_stm32_dma_tx_enable(dev);
+		}
+
+		irq_unlock(key);
+	} else {
+		/* Enable TX DMA requests */
+		uart_stm32_dma_tx_enable(dev);
+	}
 
 	return 0;
 }